Accurate railway track detection is a critical component of intelligent railway safety systems, including obstacle intrusion warning, autonomous train navigation, and infrastructure monitoring. Reliable detection improves operational safety and enables real-time decision-making in complex railway environments. However, railway track detection remains challenging due to various environmental and structural factors. Railway tracks are often affected by strong background interference caused by ballast, vegetation, shadows, and surrounding infrastructure. These background elements share similar visual characteristics with track structures, which reduces feature discriminability. In addition, railway tracks exhibit significant scale variation across image regions, ranging from large, clearly visible structures in near-view areas to thin and indistinct structures in distant regions. This variation requires the model to maintain strong multi-scale feature representation capability. Furthermore, railway scenes frequently contain complex geometric configurations, such as curved tracks, intersections, and switches, which require the model to capture long-range spatial dependencies and preserve structural continuity. Existing deep learning-based methods, particularly those derived from lane detection frameworks, improve detection performance by leveraging convolutional neural networks and structured prediction strategies. However, these methods often suffer from insufficient receptive fields, limited multi-scale context modeling capability, and weak suppression of background interference. These limitations reduce detection accuracy and robustness in challenging railway scenarios. Therefore, developing an efficient and robust railway track detection method capable of enhancing multi-scale feature representation and suppressing background interference is essential for improving intelligent railway safety systems.

A railway track detection network named dilated and masked attention network (DMA-Net) was developed to address these challenges. The proposed method was built upon a row-based anchor framework, which efficiently localizes horizontal track positions and preserves structural consistency. Several key modules were introduced to improve feature representation and discrimination capability.A dilated multi-scale fusion module (DMSFM) was designed to enhance multi-scale feature extraction. This module employed parallel dilated convolutions with different dilation rates to capture contextual information at multiple spatial scales. Dilated convolutions expanded the receptive field without reducing spatial resolution or increasing computational complexity. The parallel structure enabled effective aggregation of both local structural features and global contextual information, improving representation of curved, distant, and intersecting tracks.An adaptive multi-head masked attention (AMMA) mechanism was introduced to suppress background interference and enhance track-related feature representation. This module applied a learnable mask to the attention map, which selectively suppressed irrelevant background regions and emphasized important track areas. The multi-head attention structure enabled the model to learn diverse feature relationships and improve discrimination capability in complex environments.In addition, the SimAM attention module and DropBlock regularization were incorporated to further improve feature quality and model robustness. SimAM enhanced feature discriminability by evaluating neuron importance without introducing additional parameters. DropBlock improved generalization capability by randomly suppressing contiguous feature regions during training. These modules improved robustness while maintaining a lightweight network structure.

Experiments were conducted on the public Rail-DB benchmark dataset to evaluate the effectiveness of DMA-Net. The proposed method achieved a detection accuracy of 94.53%, which exceeded the original baseline rail detection method by 1.73%. Compared with other mainstream railway track detection methods, DMA-Net improved accuracy by an average of 8.07%. These results demonstrate the effectiveness of the proposed multi-scale fusion and masked attention mechanisms.The DMSFM module significantly improved the model’s ability to capture track features at different spatial scales. The enlarged receptive field enabled better representation of long-range contextual information, which improved detection performance in distant and curved track regions. The AMMA mechanism effectively suppressed background interference and enhanced attention to track-related features, which improved discrimination capability in complex environments containing ballast, vegetation, and structural interference.Visual analysis further demonstrated that DMA-Net produced smoother and more continuous detection results compared with existing methods. Performance improvements were particularly evident in challenging scenarios such as nighttime conditions, occlusion, and complex track intersections. These results confirm that the proposed modules effectively improve feature representation and detection robustness. DMA-Net maintained a lightweight design while achieving high detection accuracy. The proposed network achieved a processing speed of 268 f/s, which satisfies real-time processing requirements. The efficient architecture and lightweight attention mechanisms enable practical deployment in real-world railway safety systems.

DMA-Net provides an effective and robust solution for railway track detection in complex environments. The DMSFM module enhances multi-scale feature representation and enlarges the receptive field, improving the model’s ability to capture track geometry across different spatial scales. The AMMA mechanism improves discrimination capability by suppressing background interference and emphasizing track-related features. The integration of SimAM attention and DropBlock regularization further improves robustness while maintaining computational efficiency. Experimental results demonstrate that DMA-Net achieves superior detection accuracy, robustness, and real-time performance compared with existing methods. The proposed network provides a reliable technical solution for intelligent railway monitoring, autonomous train navigation, and railway safety applications.

Fluorescence-guided surgery (FGS) has emerged as a critical technique in modern medical procedures, utilizing fluorophores such as indocyanine green (ICG) to visualize anatomical structures and pathological tissues that are invisible to the naked eye. While effective, the clinical utility of fluorescence imaging relies heavily on its integration with white-light (WL) imaging to provide necessary anatomical context. Conventional dual-camera systems, which employ beam splitters and separate sensors, offer high performance but suffer from bulky form factors, high costs, and physical registration errors. Conversely, single-camera solutions that utilize time-division multiplexing often face significant technical bottlenecks. These include data congestion when processing high-bandwidth 4K video streams via software, and, more critically, the inability to balance exposure between high-dynamic-range white-light scenes and weak fluorescence signals using a single sensor. The primary objective of this study was to overcome these limitations by developing a "virtual dual-camera" architecture. This system aimed to achieve the performance and independent image processing capabilities of a physical dual-camera setup within a single-sensor hardware framework, specifically targeting real-time performance, hardware-level synchronization, and optimized automatic exposure control (AEC) for composite endoscopic imaging.

The proposed system integrated a custom hardware platform with a novel control architecture. The optical front-end utilized a high-resolution CMOS image sensor (Sony IMX334) paired with a dual-bandpass filter capable of blocking excitation light (780 nm) while transmitting visible light and fluorescence signals (above 820 nm). Illumination was provided by a synchronized dual-LED source delivering high-power white light and 780 nm near-infrared excitation.

The core innovation lay in the field-programmable gate array (FPGA)-based control logic. The FPGA functioned as the central timing master, generating precise frame synchronization signals. It coordinated the CMOS exposure window with the alternating strobe of white and near-infrared LEDs, achieving frame-level synchronization accuracy. To resolve the bottleneck of software-based frame separation, the system implemented a hardware-level video stream demultiplexing architecture. The FPGA intercepted the raw MIPI CSI-2 video stream from the sensor. Based on the current lighting state (odd frames for white light, even frames for fluorescence), the FPGA routed the image data to two independent direct memory access (DMA) channels and image signal processor (ISP) pipelines. This hardware separation allowed for the decoupling of image processing parameters; the white-light channel utilized standard color correction and gamma settings, while the fluorescence channel employed aggressive black level correction, higher gamma values for contrast enhancement, and temporal noise reduction, which relied on the frame continuity preserved by the hardware splitting.

Furthermore, a dual-modal joint automatic exposure control (AEC) algorithm was developed to address the drastic intensity disparity between the two imaging modes. The algorithm operated in two distinct phases separated by a single frame delay. In Phase 1, the system calculated the luminance of the white-light frame and adjusted the white LED intensity and the sensor’s global analog gain. Crucially, a constraint strategy prioritized the prevention of white-light overexposure. In Phase 2, utilizing the analog gain fixed during Phase 1, the system optimized the fluorescence image brightness by independently adjusting the near-infrared excitation power and applying digital gain. This decoupled approach ensured that the shared analog gain did not compromise the white-light image quality while still allowing sufficient amplification for the weaker fluorescence signal.

Experimental validation was conducted using a porcine liver model injected with ICG, imaged at varying working distances (10 mm, 25 mm, and 40 mm) to simulate clinical conditions. The hardware synchronization mechanism achieved a control precision of 16.67 ms, strictly aligning the exposure windows with the respective light sources and eliminating channel crosstalk.

The proposed AEC algorithm demonstrated superior stability and speed compared to conventional ISP-based exposure algorithms. Quantitative analysis of convergence time revealed that the system recovered to a stable brightness state within 0.8 seconds (approximately 23 frames for white light and an additional 15–20 frames for fluorescence) even under extreme lighting changes. In contrast, standard AEC algorithms exhibited significant oscillation and "flashing" artifacts due to the coupled feedback loop between the alternating bright and dark frames. The proposed method maintained the white-light luminance within a tight tolerance (50 luma) without the overshoot observed in conventional methods.

Image quality was assessed using structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR) metrics. At a standard working distance of 10 mm, the system achieved an SSIM of 0.9994 and a PSNR of 46.23 dB for white-light images, indicating negligible information loss compared to reference images. Fluorescence images at the same distance yielded an SSIM of 0.9893 and a PSNR of 42.25 dB. As the working distance increased to 40 mm, the fluorescence image quality naturally degraded due to the reliance on higher digital gain to compensate for signal attenuation, resulting in an SSIM of 0.9224. However, the system successfully maintained diagnostic visibility without introducing significant noise artifacts or false positives, a benefit attributed to the independent ISP pipeline tuning enabled by the virtual dual-camera architecture. Visual inspection confirmed that the fused images retained clear anatomical textures from the white-light channel and distinct, high-contrast tumor margins from the fluorescence channel.

The study successfully implemented a virtual dual-camera imaging system that resolves the fundamental conflicts inherent in single-sensor fluorescence endoscopy. By shifting the frame synchronization and video demultiplexing tasks to the FPGA hardware layer, the design eliminated the latency and bandwidth limitations associated with software processing, enabling smooth 4K real-time imaging. The novel two-stage automatic exposure control algorithm effectively decoupled the brightness regulation of white-light and fluorescence modes, allowing for optimal dynamic range utilization for both anatomy and pathology without physical filter switching or dual-sensor alignment.

The experimental results confirmed that this architecture delivers image quality comparable to dual-camera systems, with high structural similarity and signal-to-noise ratios, while retaining the cost and size advantages of a single-chip solution. The fast convergence time of the exposure algorithm ensures seamless transitions during intraoperative maneuvering. Consequently, this technology provides a robust, low-cost, and high-performance solution for fluorescence-guided surgery, offering surgeons reliable real-time navigation aids with precise anatomical and functional information fusion. The proposed architecture establishes a viable technical pathway for the next generation of miniaturized, multi-modal endoscopic platforms.

Corn, as one of the world’s most vital staple crops, plays an indispensable role in agricultural production, industrial applications, and economic development. In China, corn planting area and yield currently account for over 40% of total grain output, a proportion that continues to grow steadily, underscoring its strategic significance for national food security and economic stability. However, throughout the growth cycle, corn is persistently threatened by various pests and diseases, including corn borer, leaf spot, rust, stalk rot, and armyworm, which severely restrict both yield and quality. Traditional pest and disease detection methods rely heavily on manual field inspection—a labor-intensive, time-consuming, and inefficient approach that struggles to meet the demands of large-scale cultivation. Moreover, manual detection is inherently subjective and discontinuous, often resulting in missed diagnoses and misjudgments. Consequently, developing an automated, efficient, and accurate detection method for corn pests and diseases has become an urgent necessity.

In recent years, deep learning-based object detection technologies have achieved remarkable success across various domains, with the YOLO (You Only Look Once) series of algorithms gaining particular prominence due to their exceptional balance of detection speed and accuracy. YOLOv11, the latest iteration in this series, incorporates further optimizations in training strategies, network architecture, and detection precision, providing a solid technical foundation for agricultural pest identification tasks. Nevertheless, when applied to corn pest detection in real-world field scenarios, mainstream models including YOLOv11 continue to exhibit critical limitations: 1) Complex environmental interference—the intricate backgrounds, variable lighting conditions, diverse pest morphologies, and frequent occlusions characteristic of field environments substantially impair the model’s ability to accurately localize and recognize targets; 2) Small-target detection difficulty—early-stage or physically small pests and disease lesions occupy extremely limited pixel regions in images and display indistinct features, leading to elevated rates of missed detections and false negatives.

To address these challenges, this paper proposes an improved YOLOv11 model designated as MSAF-YOLO (Multi-Scale Adaptive Feedback YOLO). The core innovation lies in reconstructing the backbone network using a novel MSAF-Net (Multi-Scale Adaptive Feedback Network) architecture, which synergistically integrates two specially designed modules: the Hierarchical Feature Feedback Module (HFFM) and the Adaptive Dynamic Receptive Field Module (ADRF).

The Hierarchical Feature Feedback Module addresses the fundamental problem of semantic information deficiency in shallow feature layers and the progressive loss of small-target features during repeated downsampling operations. Conventional feedforward convolutional networks propagate information unidirectionally from shallow to deep layers; while deep layers acquire rich semantic representations, shallow layers remain dominated by fine-grained spatial details but lack high-level semantic context. HFFM introduces residual-style feedback pathways coupled with channel attention mechanisms to establish cross-layer feature calibration and enhancement. Specifically, deep feature maps containing abstracted semantic information are selectively fed back and fused with corresponding shallow feature maps through learnable gating mechanisms. The channel attention component adaptively recalibrates feature responses, emphasizing informative channels while suppressing less useful ones. This feedback-driven feature refinement significantly enhances the model’s capacity to extract and preserve discriminative characteristics of small pests and incipient disease symptoms, effectively mitigating the information attenuation typically observed across deep convolutional architectures.

The Adaptive Dynamic Receptive Field Module confronts the inherent inflexibility of standard convolutional neural networks, which employ fixed geometric receptive fields that remain constant regardless of input content. Such rigidity proves particularly disadvantageous when processing agricultural images containing objects at vastly different scales—from millimeter-scale rust pustules to centimeter-scale expanded leaf spot lesions—under heterogeneous background conditions. ADRF achieves pixel-level receptive field adaptation through an elegantly designed integration of spatially adaptive weights and multi-branch dilated convolutions. For each spatial location in the feature map, the module dynamically computes context-dependent weights that govern the relative contributions of parallel convolutional branches employing different dilation rates. This mechanism enables the network to expand its receptive field when processing large contextual regions while maintaining fine resolution for small-target localization, all modulated by the local structural characteristics of the input features. The resulting adaptive receptive field substantially improves model robustness and generalization across diverse pest and disease categories, growth stages, and environmental conditions.

The proposed MSAF-YOLO model was rigorously evaluated on two datasets: a self-constructed corn pest and disease dataset (CORN) encompassing representative categories including corn borer, leaf spot, and rust under various field conditions, and the publicly available general plant disease dataset (diseases) to assess cross-domain generalization capability. Experimental results demonstrate substantial and consistent improvements across multiple evaluation metrics. On the CORN dataset, MSAF-YOLO achieved mean Average Precision at 50% IoU (mAP50) of 85.7% and mAP50:95 of 68.9%, representing significant gains of 3.2 and 2.7 percentage points respectively over the baseline YOLOv11 model. On the diseases dataset, mAP50 and mAP50:95 reached 80.9% and 62.8%, improving by 3.7% and 3.1% respectively. Notably, these accuracy enhancements were accomplished alongside parameter optimization, demonstrating that MSAF-YOLO achieves a favorable trade-off between detection precision and computational efficiency.

In conclusion, this research makes three principal contributions: 1) it systematically identifies and characterizes the specific challenges confronting corn pest and disease detection in complex field environments, particularly the coupled difficulties of small-target detection and background interference; 2) it proposes two novel modules—HFFM and ADRF—that respectively address semantic feature deficiency and fixed receptive field limitations through innovative feedback mechanisms and dynamic convolution strategies; 3) it delivers a practically deployable model that achieves state-of-the-art detection performance on corn pest and disease tasks while maintaining architectural efficiency suitable for real-time agricultural applications. The proposed MSAF-YOLO framework not only advances the methodological frontier of object detection in agricultural contexts but also offers a viable technical pathway toward intelligent, automated crop health monitoring systems.

Infrared small target detection (IRSTD) serves as a foundational enabling technology in computer vision, with extensive and critical deployment across military and civilian domains including infrared early warning, precision guidance, maritime search and rescue, and intelligent traffic monitoring. The task presents inherent and unresolved challenges stemming from the unique physical properties of infrared small targets, which typically manifest as spot-like structures spanning only several to dozens of pixels, yielding extremely sparse texture and structural information. Compounding these difficulties, infrared imagery is frequently degraded by intense background clutter, low contrast, and low signal-to-clutter ratio (SCR), all of which obscure the faint signals of dim targets. Existing mainstream methods exhibit notable limitations in addressing these interconnected challenges: segmentation-based deep learning frameworks, while proficient in fine-grained local detail capture, incur prohibitive computational overhead and limited real-time performance; detection-based architectures, despite superior inference efficiency, demonstrate insufficient sensitivity to dim targets, suffer from target feature suppression under clutter interference, underutilize multi-scale feature information, and remain highly sensitive to minor localization deviations during training due to inherent constraints of standard intersection over union (IoU)-based loss functions. To mitigate these critical bottlenecks, this paper proposes a novel multi-scale feature learning network, hereafter referred to as MFLNet, for high-precision and efficient IRSTD.

To address the aforementioned challenges, MFLNet is built on a synergistic architecture with three core technical innovations. First, to enhance weak target feature perception in low-SCR scenarios, we develop a CSP with mamba-inspired linear attention block, termed CMABlock, for feature extraction. This module integrates cross stage partial architecture with a tailored linear attention mechanism, and incorporates rotary position encodings within a mamba-inspired linear attention submodule to realize position-aware hierarchical modeling of potential target regions. This design directs the model to focus on faint, location-critical target cues in intense background noise, improving feature discriminability while preserving a lightweight computational profile. Second, to resolve insufficient multi-scale feature utilization, we design a dynamic multi-scale feature fusion head, termed DMSFHead, as the detection decoder. Unlike conventional detection heads with fixed fusion weights, DMSFHead adopts a dynamic weighted fusion strategy, extending standard three-layer feature fusion to a four-layer feature pyramid optimized for IRSTD to retain critical high-resolution details before they are degraded in deep network layers. Its core feature aggregation and selective feature fusion modules adaptively balance contributions from multi-scale features, preserving spatial details from high-resolution layers while integrating semantic information from low-resolution layers, thus enhancing feature utilization and target-clutter discrimination. Third, to alleviate the extreme sensitivity of standard IoU metrics to minor positional shifts in tiny targets, we introduce normalized Gaussian Wasserstein distance (NWD), which models bounding boxes as two-dimensional Gaussian distributions to generate a smoother, scale-robust similarity metric. To balance this with precise geometric regression constraints, we construct a hybrid CIoU-NWD loss function, which combines the training stability of NWD for tiny targets with the accurate bounding box regression capability of complete IoU (CIoU), mitigating issues of sparse positive samples and unstable model training.

Extensive validation experiments are conducted on three widely used challenging public IRSTD datasets: IRSTD-1k, NUDT-SIRST, and NUAA-SIRST, with MFLNet benchmarked against six state-of-the-art (SOTA) methods including MDvsFA, ACM, ALCNet, ISNet, AGPCNet, and DNANet. Quantitative results show that MFLNet consistently outperforms all competing methods across all benchmarks. Specifically, MFLNet achieves an F1-score of 0.845 on IRSTD-1k, outperforming the second-best method by 10.1% relative; an F1-score of 0.945 on NUDT-SIRST, corresponding to a 4.4% relative improvement; and an F1-score of 0.889 on NUAA-SIRST, exceeding the runner-up by 4.8% relative. Ablation studies systematically validate the effectiveness of each proposed component: replacing the attention branch of CMABlock with standard convolutions incurs consistent performance degradation across all metrics, while substituting DMSFHead with a conventional detection head leads to significant drops in both precision and recall, verifying the critical contribution of dynamic multi-scale fusion. Experiments on the hybrid loss function confirm that a balanced weighting factor of λ=0.1 achieves optimal performance by reconciling training stability and localization accuracy. Cross-dataset generalization experiments further demonstrate MFLNet’s robustness to domain shifts, with stable performance across all six transfer tasks, indicating its ability to learn generalizable target representations rather than dataset-specific biases. Computational efficiency analysis shows that MFLNet maintains an exceptionally lightweight profile, with 4.12 M parameters, 13.80 GFLOPs of computational complexity, and 47.40 f/s inference speed, outperforming all compared methods in efficiency. This performance stems from its detection-based paradigm and lightweight module design, making it well-suited for deployment on resource-constrained edge platforms.

This paper proposes MFLNet, a novel high-performance multi-scale feature learning network for IRSTD. Through the synergistic design of three core components—CMABlock for enhanced weak target feature perception, DMSFHead for efficient multi-scale feature utilization, and the CIoU-NWD hybrid loss for stable model training—MFLNet systematically addresses the core technical challenges of IRSTD. Extensive experimental results confirm that MFLNet not only establishes new SOTA detection accuracy across multiple mainstream benchmarks but also exhibits superior computational efficiency. This work provides a robust and practical framework for real-world high-performance infrared detection systems, with important application value for military surveillance, autonomous navigation, and related fields. Future work will focus on optimizing sequence modeling mechanisms and exploring multi-frame spatiotemporal fusion to further improve detection performance in extremely low-SCR scenarios.

Three-dimensional lane detection provides essential geometric information for autonomous driving systems, supporting vehicle localization, trajectory planning, and lateral control. Compared with traditional two-dimensional lane detection, three-dimensional lane detection aims to recover the spatial structure of lane markings directly from monocular images. This task remains challenging because monocular imagery lacks explicit depth cues and is affected by perspective distortion, scale variation, and complex road environments. Recent studies have increasingly adopted bird's-eye-view representations that transform front-view image features into a top-down spatial layout. Such representations simplify geometric reasoning and facilitate the estimation of lane structures in three-dimensional space. However, existing convolutional neural network architectures still face difficulties in simultaneously preserving detailed lane geometry and extracting high-level semantic information. Shallow layers capture fine textures and structural boundaries, whereas deeper layers emphasize semantic abstraction and contextual understanding. Excessive dependence on deep features may introduce localization offsets after spatial projection, while shallow features alone are insufficient for reliable detection under challenging conditions such as illumination changes, occlusions, or background interference. A network architecture capable of integrating geometric detail with semantic representation is therefore important for improving monocular three-dimensional lane detection.

To address this issue, a monocular three-dimensional lane detection framework named HA-Lane Net was developed. The framework focuses on constructing balanced feature representations through a hybrid encoding strategy together with an attention-based feature refinement mechanism.

The backbone of the proposed framework is a hybrid encoder that combines complementary properties of two classical convolutional architectures. Shallow layers derived from VGG were used to capture detailed textures and structural patterns in lane regions. These features preserve spatial cues that are important for accurate geometric reconstruction. In parallel, deeper layers based on ResNet were employed to extract semantic representations and contextual information from complex road scenes. To integrate these heterogeneous features effectively, channel adaptation and hierarchical coupling mechanisms were introduced. These operations align feature dimensions and progressively fuse shallow geometric cues with deep semantic representations across multiple network stages. As a result, the encoder constructs a multi-scale feature hierarchy that maintains geometric precision while retaining strong semantic discrimination.

Feature refinement was further introduced before spatial projection through a pre-projection channel attention mechanism. Inspired by the squeeze-and-excitation principle, this module recalibrates feature responses by strengthening channels related to lane structures and suppressing background signals. Applying attention prior to the front-view-to-bird's-eye-view projection improves feature consistency across viewpoints and reduces the influence of image blur, lighting variation, and environmental clutter. This design helps preserve structural lane information during the spatial transformation process.

Training stability was improved using a cosine annealing learning-rate schedule, which facilitates smoother optimization of the multi-task loss function. The framework was trained and evaluated on the Apollo-Sim3D dataset, which contains diverse simulated driving scenes with accurate three-dimensional lane annotations. During inference, the network directly predicts spatial coordinates and shapes of lane markings from monocular images, enabling reconstruction of lane geometry in three-dimensional space without relying on additional sensors.

Experiments on the Apollo-Sim3D dataset show that HA-Lane Net improves three-dimensional lane detection performance compared with representative monocular approaches such as 3DLaneNet and Gen-Lane-Net. The hybrid encoder effectively integrates geometric detail with semantic context, enabling more accurate estimation of lane geometry.

The pre-projection attention mechanism further improves robustness by highlighting lane-related feature responses and suppressing irrelevant background information. This targeted refinement enhances feature consistency during the transformation from front-view images to bird's-eye-view representations and reduces localization errors caused by occlusion, illumination variation, or scene complexity. Consequently, the network produces more stable detection results across different driving scenarios.

Quantitative comparisons indicate improved accuracy in three-dimensional lane coordinate estimation relative to the reference methods used for comparison. Detection results also show increased stability across diverse road environments. These improvements arise from the combined effect of hybrid feature encoding and attention-based feature refinement, which strengthen lane structure representation while mitigating background noise. In addition, the proposed framework maintains a relatively efficient architecture. Performance improvements are achieved without introducing Transformer-based components or other computationally intensive modules, allowing the network to retain manageable complexity for real-time perception tasks.

A monocular three-dimensional lane detection framework named HA-Lane Net has been presented for improved lane perception in autonomous driving environments. The proposed approach integrates a hybrid encoder that combines shallow VGG layers with deep ResNet layers, enabling balanced extraction of geometric detail and semantic representation. A pre-projection channel attention mechanism further refines lane-related features before spatial transformation, improving robustness against environmental interference.

Experimental evaluation on the Apollo-Sim3D dataset shows that the proposed framework improves three-dimensional lane detection performance compared with representative monocular approaches such as 3DLaneNet and Gen-Lane-Net. The hybrid encoding strategy together with attention-based refinement enables more reliable estimation of lane geometry while maintaining a relatively efficient network structure. These results suggest that the proposed framework provides a practical solution for monocular three-dimensional lane perception and offers useful insights for future research on autonomous driving scene understanding.

Side-scan sonar (SSS) is an indispensable acoustic imaging technology for seabed exploration, underwater archeology, and search and rescue operations. It provides wide-range, high-resolution images of the seafloor by transmitting acoustic pulses and recording their backscattered echoes. However, due to the complex underwater environment and the inherent characteristics of acoustic imaging, SSS images are typically contaminated by severe speckle noise, geometric distortions, uneven resolution, and blurred target boundaries. These degradations significantly impair the performance of conventional object detection models that are originally designed for optical images. Furthermore, SSS datasets are often limited in scale and suffer from inconsistent annotation quality, as the low resolution and fuzzy features make it difficult to label targets precisely. Such low-quality training samples introduce harmful gradients during optimization, further compromising model robustness. To address these challenges, this paper proposes an improved detection model named YOLOv8-Efficient Sonar Image (YOLOv8-ESI), aiming to achieve accurate, robust, and lightweight underwater target detection that meets the real-time requirements of resource-constrained autonomous underwater vehicles.

YOLOv8-ESI is developed based on the YOLOv8n baseline through three synergistic innovations tailored to SSS image characteristics. First, a wavelet convolutional layer (WTConv) is integrated into the backbone network. Unlike standard convolutions that operate solely in the spatial domain, WTConv performs cascaded two-dimensional Haar wavelet decomposition, transforming feature extraction into the frequency domain. The low-frequency sub-band (LL) captures global structural information such as the highlight-shadow patterns of sonar targets, while the high-frequency sub-bands (LH, HL, HH) preserve edge details and textures. By applying small-sized depthwise convolutions to these sub-bands, WTConv effectively enlarges the receptive field without introducing excessive parameters. Moreover, the multi-scale wavelet decomposition enables adaptive noise suppression through thresholding in the frequency domain, enhancing the model's ability to distinguish true targets from background clutter. Second, the neck network is systematically reconstructed into a dynamic feature pyramid network (D-GFPN). This architecture incorporates an efficient dynamic upsampler (DySample) that employs a multi-level point sampling strategy to preserve fine-grained details during feature resolution recovery, which is critical for small target detection in SSS images. A cross-stage partial dual-path lightweight aggregation network module (CSP-DLAN) is also designed, utilizing parallel 1×1 and 3×3 grouped depthwise convolutions along with channel shuffle mechanisms to achieve efficient multi-scale feature fusion with minimal computational overhead. This design balances the extraction of spatial details and channel-wise interactions while promoting cross-group information flow. Third, the original CIoU loss is replaced with WIoU, which employs a dynamic non-monotonic focusing mechanism based on outlier degree to assess anchor box quality. By assigning smaller gradient gains to low-quality samples and focusing on medium-quality anchors, WIoU effectively mitigates the harmful effects of annotation noise and bounding box inaccuracies prevalent in SSS datasets, thereby improving training stability and overall detection accuracy.

Experiments are conducted on a self-constructed SSS dataset (SSSD) that integrates publicly available sources (Seabed Objects-KLSG and SONAR-2019) with additional images collected from underwater experiments and online repositories. The dataset contains 956 original images of three target categories: shipwrecks (602 images), aircraft wrecks (279 images), and human remains (75 images), with resolutions ranging from 300 pixel×300 pixel to 800 pixel×500 pixel. To enhance generalization and prevent overfitting, the training set is augmented through horizontal flipping, random rotation, random brightness adjustment, and Gaussian noise addition, resulting in 2926 training samples. The dataset is split into training, validation, and test sets with a ratio of 7:2:1. Ablation studies are performed to validate the contribution of each proposed module. Integrating WTConv alone improves mAP50 by 2.6% and mAP50:95 by 2.5% while reducing parameters, demonstrating its effectiveness in frequency-domain feature enhancement and noise suppression. The proposed D-GFPN outperforms mainstream neck structures including PANet, BiFPN, HS-FPN, and RepGFPN, achieving 87.3% mAP50 with only 2.98 M parameters and 7.8 GFLOPs, representing a 9.1% parameter reduction and 8.3% FLOPs reduction compared to RepGFPN while maintaining the same mAP50. Among different WIoU variants, WIoUv3 yields the best performance, particularly improving detection of challenging categories such as human remains by 1.7%, which validates its effectiveness in handling low-quality samples. The complete YOLOv8-ESI model achieves 88.8% mAP50, and 63.4% mAP50:95, surpassing baseline YOLOv8n by 5.2% and 3.8%, respectively, while reducing parameters by 10% (from 3.01 M to 2.69 M) and computational complexity by 1.1 GFLOPs. Comparative experiments against state-of-the-art detectors including YOLOv3-tiny, YOLOv5s, YOLOv6n, YOLOv9c, YOLOv10n, YOLOv11n, RT-DETR, SCR-YOLOv8, and CSC-YOLO demonstrate that YOLOv8-ESI achieves superior detection accuracy with competitive inference speed (1.3 ms per image) and model compactness. Heatmap visualizations confirm that the model focuses precisely on target regions with suppressed background activation, while detection examples illustrate reduced false positives and improved boundary localization compared to baseline models.

This paper presents YOLOv8-ESI, a lightweight and accurate detection model specifically designed for side-scan sonar imagery. By integrating wavelet convolution for frequency-domain feature enhancement, a dynamic feature pyramid network for adaptive multi-scale fusion, and a wise IoU loss for robust training against annotation noise, the model effectively addresses the core challenges of noise interference, scale variability, feature blurring, and low-quality annotations inherent in SSS data. Experimental results demonstrate that YOLOv8-ESI significantly outperforms baseline and state-of-the-art models in both accuracy and efficiency, meeting the real-time detection requirements of underwater platforms. The ablation studies validate the synergistic effects of the proposed modules, and the comparative experiments confirm the model's superiority across multiple evaluation metrics. This work provides an effective technical solution for advancing underwater object detection in marine exploration, archaeology, and search and rescue operations. Future work may explore the integration of advanced denoising techniques and the extension of the model to other underwater imaging modalities such as synthetic aperture sonar and multi-beam echo sounders.

Image classification serves as a fundamental task in computer vision, aiming to automatically assign semantic labels to input images through algorithmic models. Although deep learning, particularly convolutional neural networks (CNNs), has achieved remarkable progress in this domain, existing approaches still face significant challenges. Traditional attention mechanisms often overemphasize global relationships while sacrificing the perception of local fine-grained features and textures. Furthermore, these methods struggle to adapt to variations in object scale and spatial relationships, failing to fully integrate global spatial contextual information. The residual connections in deep networks can amplify noise and adversarial perturbations, leading to feature dilution in identity mappings. Additionally, most current architectures lack dynamic weight assignment capabilities, resulting in limited interpretability and flexible adaptability. To address these limitations, this paper proposes a Spatial Dynamic Perception Image Classification Network (SDPNet) that enhances the dynamic perception of critical features, strengthens spatial-channel interactions, improves robustness to geometric transformations, and increases overall network generalization capability.

The proposed SDPNet is constructed upon the ResNet-34 architecture with three key innovations. First, the Negative Exponent Suppression (NES) module is introduced as a novel nonlinear feature transformation technique. This module applies an absolute value transformation to remove polarity differences while preserving feature continuity, followed by an exponential decay suppression operation that replaces negative values with their absolute values multiplied by a decay factor of exp(-6), compressing negative responses to approximately 0.25% of their original magnitude. This approach enhances feature contrast, highlights edge information, and strengthens the network's ability to learn discriminative features while maintaining a balance between information preservation and discriminative enhancement. Second, the Spatial Dynamic Perception (SDP) module is developed, integrating the NES module with three parallel processing branches: Feature Rotation (FR), Dynamic Expansion (DE), and Channel Shuffle (CS). The FR branch performs 180-degree center-symmetric rotation of feature maps followed by 1×1 depthwise separable convolutions and Tanh activation, capturing spatial symmetry and improving robustness to object orientation changes. The DE branch employs convolutional and deconvolutional operations with 3×3 and 2×2 kernels respectively, dynamically expanding receptive fields while preserving spatial resolution through nearest-neighbor interpolation upsampling, thereby capturing both fine details and holistic contours. The CS branch utilizes two 1×1 convolutions with channel shuffling between them to break fixed channel connection patterns, enhancing feature diversity and enabling cross-channel information exchange. These three branches are fused through spatial and channel average pooling followed by Tanh activation, enabling adaptive feature enhancement and dynamic selection of discriminative spatial information. Third, the Dual-Dimension Fusion Residual (DFR) module is proposed, which embeds the SDP attention mechanism into residual blocks and incorporates conditional convolutions into the dimension-reduced residual branches. The conditional convolution dynamically adjusts kernel weights based on input features through a routing function, transforming static convolution into dynamic, condition-dependent operations. This design enhances spatial and channel interactions, optimizes feature representation, reduces overfitting, and improves generalization capability. Additionally, the initial feature extraction layer is modified by replacing the 7×7 convolutional kernel with a 3×3 kernel and removing the max pooling layer to better preserve local details and original spatial information.

Comprehensive experiments were conducted on five benchmark datasets: CIFAR-100, CIFAR-10, SVHN, Flowers-102, and NWPU-RESISC45. SDPNet achieved classification accuracies of 80.63%, 96.95%, 97.66%, 90.34%, and 92.46% respectively, outperforming eleven state-of-the-art models including ResNet-34, EfficientNets, GhostNet, QKFormer, TLENet, ATONet, Couplformer, SSLLNet, Multi-ResNet, DCDENet, and MobileNet-LAM. Compared with these competing methods, SDPNet demonstrated average accuracy improvements of 3.60%, 2.52%, 2.09%, 7.29%, and 2.36% across the respective datasets. The model's parameter count (21.93 M) remains comparable to ResNet-34 (21.35 M) while being substantially lower than EfficientNets (53.68 M) and Multi-ResNet (52.03 M). The computational complexity (7.5 GFLOPs) and inference speed (2.55 f/s) demonstrate practical deployability in resource-constrained scenarios. Ablation studies confirm the complementary contributions of each module: SDP alone improves ResNet-34 by up to 4.14% on Flowers-102, DFR alone improves by up to 2.65%, and their combination yields synergistic gains of up to 7.59%, validating the effectiveness of the integrated design. Attention mechanism comparisons against SE, SCConv, CA, CBAM, and ADWM across CIFAR-100, Flowers-102, and STL-10 datasets show that SDP consistently achieves superior or comparable performance, with average improvements of 0.6%, 1.27%, and 1.56% respectively. The 180° rotation angle in the FR branch was empirically determined as optimal through systematic parameter analysis. Experiments on SDP module combinations reveal that the full NES+FR+DE+CS configuration achieves the highest accuracy, with NES identified as the core effective component. Position sensitivity analysis indicates that placing SDP at deeper network layers (position 1) maximizes its effectiveness by leveraging larger receptive fields for hierarchical spatial-channel interaction capture. Optimal DFR module placement involves embedding three D-Blocks within the residual connections. Visualization analysis through heatmaps demonstrates that SDPNet effectively suppresses noise and redundant information while focusing on critical regions, exhibiting more refined and comprehensive feature extraction capabilities compared to MobileNet_V2, EfficientNet, and ResNet-34.

This paper presents SDPNet, a novel image classification network that addresses fundamental limitations in existing approaches through three synergistic contributions. The NES module enhances spatial feature perception through nonlinear feature transformation, the SDP module dynamically captures discriminative spatial information through integrated rotation, expansion, and shuffling operations, and the DFR module optimizes spatial-channel interactions through conditional convolutions. The comprehensive experimental validation across multiple datasets demonstrates SDPNet's superior classification performance, robust generalization capability, and practical efficiency. Future work will extend this spatial-domain approach by incorporating frequency-domain perception to further enhance classification performance.

The three-dimensional structure of oceanographic parameter profiles, particularly temperature and salinity, has persistently posed a significant challenge in remote sensing detection. While traditional ocean color satellites and microwave radiometers can achieve sea surface temperature measurements with accuracies better than 0.5 K and salinity measurements with monthly averages of 0.1 psu, these techniques are fundamentally limited to two-dimensional planar observations. In situ measurements from ships and buoys, though capable of vertical profiling, yield discrete, asynchronous, and non-global data that cannot satisfy the requirements of ocean science, environmental monitoring, marine economy, and national defense. Conventional oceanographic lidar systems, which rely solely on echo intensity detection, are unable to simultaneously measure multiple parameters like temperature and salinity. To address this critical gap, this paper proposes a novel remote sensing system for multi-parameter oceanographic detection based on Brillouin laser spectroscopy.

The proposed system leverages the laser non-elastic scattering effect, specifically spontaneous Brillouin scattering, which occurs when photons interact with thermal acoustic phonons in seawater. The Brillouin frequency shift and linewidth are functions of seawater temperature and salinity, as they depend on physical properties including sound velocity, refractive index, density, and viscosity coefficients. A two-parameter inversion model is established by solving coupled equations that relate Brillouin shift and linewidth to temperature and salinity. Through regression fitting of numerical data generated across temperature ranges of 0–30 °C and salinity ranges of 0–35 ‰, empirical formulas are derived to directly calculate temperature and salinity from measured Brillouin parameters.

The detection system comprises a transmitter system, a receiver system, and an integrated control and data processing system. A frequency-stabilized narrow-linewidth laser employing injection seeding and iodine molecular frequency stabilization technology serves as the light source, with a center frequency stability of 6 MHz, wavelength output at 532.29 nm, single pulse energy of 1 mJ, and pulse width of 7.6 ns. The core spectroscopic detection module consists of a high-precision Fizeau interferometer coupled with a highly sensitive 16-channel photomultiplier tube (PMT) array. The Fizeau interferometer is designed with a plate spacing of 6 mm to achieve a free spectral range of 25 GHz, a mirror reflectivity of 0.873 to yield an instrumental function linewidth of 1.079 GHz, and a plate angle of 6.65 μrad with a clear aperture of 40 mm. When the laser beam passes through the Fizeau interferometer, the Rayleigh and two Brillouin components form three distinct transmission peaks at different positions on the PMT array. The multi-channel signals are processed through baseline correction, abnormal signal rejection, noise suppression, and cumulative averaging. A nonlinear least-squares fitting algorithm is then employed to reconstruct the discrete Brillouin spectrum, extracting the frequency shift and linewidth from the measured data.

Simulation analyses reveal that to achieve temperature retrieval accuracy ≤1 K and salinity accuracy ≤2 psu for typical seawater conditions (temperature 10-25 °C, salinity 30-35 psu), the measurement uncertainty of both Brillouin shift and linewidth must be controlled within 20 MHz. This requirement drives the stringent design specifications of the system components. The PMT channel number selection is optimized based on simulation results showing that 8 or more channels satisfy the detection requirements, with 16 channels selected for the actual system implementation.

Laboratory experiments were conducted in a controlled water tank environment using two separate tanks to simulate layered water body slices, with independently controllable temperature and salinity conditions. Tank 1 was maintained at 23.1 °C and 33.8 psu, while Tank 2 was maintained at 23.4 °C and 34.8 psu. The laser pulses sequentially passed through both tanks, and the backscattered signals were received and processed. After data accumulation of 8000 pulses, the system achieved temperature measurement accuracy of 0.37 °C and salinity accuracy of 0.79 psu, with maximum errors of 0.72 °C and 1.52 psu respectively over five repeated measurements. The relationship between accumulation counts and measurement accuracy was investigated, showing that accuracy improves with increasing accumulation counts and converges to the system's optimal performance at approximately 8000 accumulations.

This paper successfully demonstrates a Brillouin spectroscopy-based remote sensing system capable of simultaneous temperature and salinity profiling in seawater. The system combines a frequency-stabilized laser, a custom-designed Fizeau interferometer with optimized parameters, and a multi-channel PMT array for high-resolution spectral detection. The two-parameter inversion model effectively derives temperature and salinity from measured Brillouin shift and linewidth. Laboratory validation confirms that the system achieves temperature accuracy of 0.37 °C and salinity accuracy of 0.79 psu, exceeding the design targets of 1 K and 2 psu respectively. These results establish the feasibility of using Brillouin lidar for remote sensing of oceanographic vertical profiles, providing a promising new approach to address the long-standing challenge of three-dimensional ocean parameter measurement. Future work will focus on field deployment and validation under real oceanic conditions.

Submillimeter wave telescopes require reflector panel surface accuracy at the micrometer level. This precision is essential for diffraction-limited performance at wavelengths near 350 μm. Panel support structures must resolve a fundamental engineering conflict. They need high static stiffness to resist gravity-induced deformation under varying elevation angles. At the same time, they require controlled flexibility to accommodate differential thermal expansion without generating stress concentrations or warping. Traditional design approaches rely heavily on iterative finite element analysis. These methods treat the structure as a numerical black box. They offer little insight into the physical origins of parasitic displacements and rotations. Their computational cost grows rapidly with model complexity. Design cycles become prohibitively slow for large-scale systems. This work addressed these limitations by developing a physics-informed analytical framework. The goal was efficient multi-objective optimization of support geometry. The framework needed to balance static rigidity and thermal adaptability in a transparent, interpretable manner.

An analytical flexibility matrix model was constructed using screw theory. The model targeted the four-point symmetric support configuration commonly used in large submillimeter telescopes. Each support leg consisted of a circular-section flexible rod. These rods acted as compliant elements with well-defined mechanical behavior. Closed-form displacement-force relationships were derived for individual rods. The same formulation extended to the integrated panel-support system. These expressions enabled direct calculation of panel-level rigid-body motions and elastic deformations from applied loads. No numerical solvers were required during evaluation. All analytical flexibility coefficients underwent rigorous validation. High-fidelity finite element models served as benchmarks. Absolute relative errors across all coefficients remained below 1.5%. A global sensitivity analysis examined entries in the flexibility and stiffness matrices. It identified specific off-diagonal terms most responsible for parasitic translations and rotations. These terms directly contributed to surface RMS error. They became the core objective functions in a multi-objective optimization problem. Design variables included the length and radius of both inner and outer support rods. These parameters control directional compliance without altering topology. The NSGA-II evolutionary algorithm generated a Pareto front of optimal solutions. The entire optimization loop operated exclusively on the analytical model. No finite element simulations were called during the process. This eliminated the primary bottleneck in conventional workflows. Preliminary experimental validation tested a full-scale aluminum panel prototype. A coordinate measuring machine recorded three-dimensional surface coordinates. Complementary photogrammetry provided dense point-cloud data. Tests covered five telescope elevation angles to simulate gravity vector changes. Thermal cycling spanned –40 °C to +40 °C in 10 °C steps. Surface repeatability and thermal drift were quantified under realistic operating conditions.

The optimized support geometry achieved synergistic static and thermal performance. Under gravity loading at zenith, axial (Z-direction) RMS surface error decreased from 7.6 μm to 7.0 μm. This 8% reduction confirmed enhanced resistance to panel sag. It also reflected effective suppression of parasitic rotation about transverse axes—key contributors to low-order surface aberrations. Circumferential (Y-direction) error remained stable at 1.4 μm across all orientations. This consistency verified preserved torsional rigidity around the optical axis. Radial (X-direction) error increased modestly from 4.9 μm to 6.0 μm. This change was not a failure but a strategic trade-off. Increased radial compliance allowed the panel to expand freely during heating. The benefit became evident under thermal load. A uniform 20 °C temperature rise reduced RMS surface error from 8.8 μm to 3.7 μm—a 58% improvement. The result demonstrated successful management of thermally induced deformation. Flexibility was deliberately distributed to align with expected thermal expansion modes. Computational efficiency improved by two orders of magnitude. The analytical optimization completed in minutes. Conventional finite element-based iteration required hours for comparable resolution. Total computing time dropped to 0.33% of the baseline—a 99.67% reduction. Experimental measurements showed surface errors stabilized between 12 μm and 13 μm. Errors varied by less than 1 μm across all five orientations. Thermal hysteresis was negligible after three cycles. Measured values matched analytical predictions within combined uncertainties. These uncertainties included initial manufacturing tolerances (±5 μm) and instrument resolution (±2 μm). The agreement validated the model’s predictive capability in real-world conditions. The controlled increase in X-direction compliance exemplifies a paradigm shift. Instead of maximizing global stiffness, the design redistributes directional compliance. This enables true multi-physics co-optimization. Parasitic motions are not eliminated—they are guided along benign paths. The approach recognizes that deformation is inevitable; the goal is to make it predictable and non-degrading to optical performance.

The screw theory-based analytical flexibility matrix model provides a transparent, interpretable foundation for high-precision support design. It establishes direct, causal links between geometric parameters and parasitic motions. This overcomes the opacity of simulation-driven black-box methods. The resulting optimization framework successfully reconciles competing demands for static rigidity and thermal compliance. It delivers measurable performance gains under both gravitational and thermal disturbances. Computational cost drops by over 99% without sacrificing accuracy. The methodology offers a robust, scalable solution for submillimeter-wave telescope reflector panels. Its principles apply broadly to other demanding opto-mechanical systems. Examples include space telescopes, cryogenic spectrometers, and adaptive secondary mirrors. Any application requiring micron-level stability under multi-physics loads can adopt this approach. The model’s analytical nature enables rapid design exploration, sensitivity studies, and real-time compensation strategies. Future work will extend the framework to asymmetric support layouts and non-circular rod cross-sections. Integration into full-reflector co-design pipelines is underway. Coupling with active alignment systems could enable closed-loop thermal compensation. This work establishes a new design philosophy: precision through intelligent flexibility, not brute-force stiffness.

Large-scale low earth orbit (LEO) satellite networks can provide strong coverage, high data transmission rates and low latency. This makes them an important part of future 6G communication systems. Routing technology is a critical component of large-scale LEO satellite networks, providing high-quality and low-latency services for network communications through packet transmission mechanisms. However, routing technology for large-scale LEO satellite networks faces critical challenges, including poor routing transmission performance due to dynamic changes in network topology and insufficient scalability of algorithmic models. In recent years, the development of artificial intelligence (AI) has provided new avenues for addressing these challenges. To address these challenges, this study utilizes AI technology, combining graph neural networks (GNNs) and deep reinforcement learning (DRL) to propose a routing algorithm that enhances transmission performance across diverse network topologies while demonstrating robust topological generalization capabilities.

This study utilized message passing neural networks (MPNNs) from GNN models and deep Q-networks (DQNs) from DRL models, and proposed a large-scale LEO satellite network routing algorithm based on MPNN and DQN (MDSR). Firstly, the MDSR algorithm employed a virtual topology approach to partition large-scale LEO satellite networks into fixed time intervals and construct them as graph-structured data for input into MDSR. Secondly, we designed the MDSR framework. The MDSR framework included two parts: an improved MPNN module and a DQN module. The improved MPNN module included a message processing stage and a readout stage. The message processing stage consisted of a message processing function (MPF), an information aggregation function (IAF), and a hidden state update function (HSUF). The readout stage consisted of a readout function (RF). Specifically, we improved the HSUF by replacing the original recurrent neural networks (RNNs) with long short-term memory (LSTM) networks. This aimed to enhance the model's ability to learn long-range dependencies and mitigate vanishing and exploding gradients. The improved MPNN module extracted the satellites and inter satellite links state features and their interconnections from the networks. The DQN module then used these features to make routing decisions. Finally, in order to compare the routing and transmission performance of different algorithms in different satellite network configurations, this study selected the HSUF as the RNN for the MPNN + DQN (HSUF-RNN) based algorithms, the Dijkstra shortest path (DSP) algorithm, the DQN based algorithm, the Q-learning-based algorithm, and the deep deterministic policy gradient (DDPG)-based algorithm for comparison with MDSR. MDSR was trained alongside other reinforcement learning algorithms used for comparison in a routing scenario involving a large-scale LEO satellite network in a polar orbit with a routing update interval of 30 s. Simultaneously, MDSR and other comparison algorithms underwent routing transmission performance testing under 10 s, 30 s, 60 s and 300 s routing update intervals for large-scale LEO satellite networks in polar and inclined orbits. The experimental environment included the Windows 11 operating system, an NVIDIA GeForce RTX 4050 GPU, the PyTorch 1.13.1 + CUDA 11.6 deep learning framework, and the Python 3.9 programming language.

The results demonstrate that, during training, MDSR achieved faster convergence and a higher initial reward than HSUF-RNN. During testing, in polar large-scale LEO satellite networks, MDSR achieved an average packet loss rate approximately 48% lower than those of the other RL algorithms and 35% lower than that of DSP. The average end-to-end delay of MDSR was approximately 20.5%, 7.9%, 2.7%, and 5.4% lower than that of DSP, Q-learning, DQN, and DDPG, respectively. The average throughput of MDSR was approximately 21.7%, 7.6%, 5%, and 1.2% higher than that of DSP, Q-learning, DQN, and DDPG, respectively. In inclined large-scale LEO satellite networks, MDSR adapted to different constellation network configurations without requiring additional training, demonstrating strong topological generalization capabilities. MDSR significantly outperformed static routing algorithms like DSP in terms of routing transmission performance. Compared to routing algorithms based on non-GNN+DRL paradigms, MDSR can accurately perceive the global network state to make routing decisions, thereby improving key routing metrics such as end-to-end delay, throughput, and packet loss rates. Therefore, in our experiments, MDSR enhanced routing and transmission performance. Without additional training, the MDSR algorithm demonstrated stable routing performance across different constellation network configurations, exhibiting strong topological generalization capabilities.

This study proposes the MDSR algorithm, which has been developed to address the issues of degraded transmission performance and insufficient topological generalization capability in routing algorithms. MDSR utilizes the improved MPNN module to extract network state features, and the DQN module learns the best routing strategies through these features. MDSR is subjected to experimental evaluation in both polar and inclined orbit large-scale LEO satellite networks. It is also compared with the HSUF-RNN, DSP, DQN, Q-learning and DDPG algorithms. The results show that, during training, MDSR exhibits faster convergence and higher initial rewards compared to the HSUF-RNN model. During testing, MDSR outperforms other comparison algorithms in both routing transmission performance metrics and topological generalization capabilities. In future research, we will explore optimization schemes such as random link connections and traffic partitioning to further enhance the robustness of MDSR while also achieving more balanced network load distribution.

Silicon-based photodetectors have established themselves as indispensable components in modern integrated optoelectronics, offering compelling advantages, including low fabrication costs and seamless compatibility with complementary metal-oxide-semiconductor (CMOS) processing infrastructure. These attributes have enabled their widespread deployment across consumer electronics, data center optical interconnects, and machine vision systems. However, silicon’s fundamental physical limitation imposes a strict cutoff wavelength of approximately 1.1 μm, rendering it inherently insensitive to the technologically critical infrared communication windows centered at 1.31 μm and 1.55 μm, as well as the extended short-wave infrared (SWIR) spectrum beyond 2 μm. This intrinsic “bandgap bottleneck” fundamentally constrains silicon photonics from fully addressing the escalating demands of next-generation applications, including fifth-generation wireless communications, light detection and ranging for autonomous vehicles, artificial intelligence-enhanced vision systems, night-vision surveillance, and environmental monitoring. Although traditional alternatives, such as germanium-based or III–V compound-semiconductor detectors, offer extended infrared response, they present significant drawbacks, including incompatibility with CMOS processes, high material costs, and integration complexity. Against this technological landscape, the emergence of deep-level impurity hyperdoping represents a promising approach to overcoming silicon’s spectral limitations while preserving its manufacturing advantages. This review provides a comprehensive and systematic examination of this transformative materials platform, offering critical insights into the fundamental mechanisms, fabrication methodologies, material characteristics, and device performance metrics that define the current state and future trajectory of hyperdoped silicon-based infrared photodetection technology.

The fabrication of hyperdoped silicon materials fundamentally relies on two complementary technological pillars: ion implantation for supersaturated impurity incorporation and subsequent ultrafast thermal processing for lattice recrystallization and electrical activation. Ion implantation enables precise control over dopant species, concentration profiles, and penetration depths, facilitating the introduction of deep-level impurities at concentrations far exceeding equilibrium solubility limits. The choice of impurity species critically determines the optoelectronic properties of hyperdoped silicon, with two principal categories demonstrating particular promise. Chalcogen elements introduce deep-level donor states that form well-defined intermediate bands (IBs) within silicon’s bandgap, enabling sub-bandgap infrared absorption. Transition metals similarly create deep-level states, though their electronic configurations and resulting band structures differ substantially from those of chalcogen-doped systems. The ultrafast annealing step, typically implemented through either pulsed laser melting (PLM) or flash lamp annealing (FLA), enables impurity redistribution and substitutional incorporation during liquid-phase epitaxial regrowth, and suppresses thermal diffusion that would otherwise lead to impurity segregation or precipitate formation. PLM offers nanosecond-scale processing with precise spatial control and a minimal thermal budget, while FLA provides larger-area uniformity and compatibility with batch processing. Both techniques successfully preserve the supersaturated impurity concentrations essential for strong sub-bandgap absorption while restoring crystalline quality and minimizing residual implantation damage.

The IB formed by deep-level impurities fundamentally alters silicon’s optical properties, introducing sub-bandgap absorption coefficients reaching 10⁴ cm⁻¹ in the near-infrared and extending sensitivity well beyond 2 μm. Photodetectors fabricated from hyperdoped silicon materials have demonstrated impressive room-temperature performance. Te-hyperdoped silicon photodiodes achieve specific detectivities of 3.2 × 10¹² cm·Hz^1/2·W⁻¹ at 1.0 μm and 9.2 × 10⁸ cm·Hz^1/2·W⁻¹ at 1.55 μm, with response times of 39-42 μs. Ag-hyperdoped silicon detectors exhibit responsivities of 504 mA/W at 1310 nm and 65 mA/W at 1550 nm under −3 V bias, with external quantum efficiencies of 0.48 and 0.052 respectively, and response times of approximately 12-16 μs. Titanium-hyperdoped devices demonstrate a responsivity of 34 mA/W at 1.55 μm with a specific detectivity of 1.7 × 10⁴ cm·Hz^1/2·W⁻¹ at 660 Hz. Au-hyperdoped silicon extends room-temperature photoresponse to 2.2 μm. Critically, these devices maintain full compatibility with standard CMOS fabrication processes, suggesting straightforward pathways toward monolithic integration. The working mechanism involves IB-assisted sub-bandgap absorption followed by gain generation through carrier-lifetime disparity; that is, photogenerated minority carriers are temporarily trapped by deep-level states, which extends their effective lifetime far beyond the electrode transit time and enables multiple majority-carrier injections per absorbed photon, providing the fundamental origin of high photoconductive gain. However, ion implantation-induced lattice defects lead to elevated dark current densities, and external quantum efficiency remains substantially below commercial detectors, fundamentally limiting sensitivity in photon-starved applications. The thermal stability of the IB also remains a concern, as high-temperature processing can lead to impurity precipitation and degradation of optoelectronic properties.

Deep-level impurity hyperdoping via ion implantation and ultrafast thermal processing has emerged as a transformative materials platform that fundamentally addresses silicon’s intrinsic bandgap limitation while preserving its manufacturing advantages. The IB formed by supersaturated chalcogen or transition metal impurities enables efficient sub-bandgap photon absorption, effectively extending silicon-based photodetector sensitivity into technologically critical infrared spectral regions. The combination of precise ion implantation with nanosecond-scale PLM or millisecond-scale FLA successfully achieves and maintains the nonequilibrium impurity concentrations essential for IB formation. To advance practical applications, future research should pursue a systematic path encompassing device architecture optimization, material system engineering, and dimensional innovation. First, device architecture optimization represents the foundational path toward performance breakthroughs. The design of CMOS-compatible lateral p-i-n photodiode arrays with ultrathin hyperdoped active layers and interdigitated electrodes can enhance carrier collection efficiency. Integration of wide-bandgap blocking layers and surface passivation coatings can effectively suppress dark current by orders of magnitude. Second, material system engineering through heterojunction integration with two-dimensional materials offers new pathways for performance enhancement. Combining hyperdoped silicon with transition-metal dichalcogenides or graphene enables band structure engineering and interfacial-field-driven carrier separation. However, lattice mismatch and interface state issues must be systematically addressed through optimized passivation strategies. Notably, recent studies demonstrate that even without hyperdoping, graphene integrated with microstructured silicon can achieve broadband infrared response through potential fluctuation engineering, highlighting alternative pathways that complement the hyperdoping approach. Third, extending hyperdoping from bulk silicon to low-dimensional nanostructures opens new frontiers for exploring quantum confinement effects. Silicon nanowires and nanomembranes offer exceptional light confinement, shortened carrier collection paths, and potential for monolithic integration with on-chip photonic circuits, though challenges in conformal doping and thermal budget management for non-planar structures remain unresolved. As these challenges are systematically addressed through coordinated experimental and theoretical efforts, deep-level impurity-hyperdoped silicon is poised to emerge as a foundational technology for next-generation infrared optoelectronics, enabling cost-effective, CMOS-compatible, room-temperature infrared detection for communications, sensing, imaging, and security applications.

Long-wave infrared (LWIR, 6-18 μm), as the core band of the atmospheric window, holds significant potential for applications in fields such as infrared imaging, target detection, solar photothermal power generation, and electromagnetic stealth. Traditional infrared absorbing materials, limited by the intrinsic electromagnetic properties of natural materials, generally face technical bottlenecks including narrow bandwidth, large thickness, and angle sensitivity. As an artificial composite material with subwavelength periodicity, metamaterials can break the limit of electromagnetic regulation of natural materials by precisely designing the equivalent permittivity and permeability of the unit structure, thus paving a new way for broadband infrared absorption technology. Researchers have gradually expanded the absorption bandwidth and performance of metamaterial absorbers through strategies such as coupling of multi-resonator units, design of composite dielectric layers, and optimization of topological structures. Nevertheless, the design of broadband metamaterial absorbers still faces three core challenges: first, most design schemes rely on the planar integration of multi-sized resonant units (such as multi-ring and multi-disk structures), and such horizontal integration tends to increase the unit cell size, which is unfavorable for device miniaturization; second, although multi-layer structures can improve absorptivity and broaden the bandwidth to a certain extent through resonant coupling, an excessive number of structural layers complicates the device architecture, hindering its engineering applications; third, the fabrication of some high-performance devices depends on noble metals (e.g., gold and silver), resulting in high manufacturing costs and poor scalability, which limits their large-scale engineering applications. Therefore, this work aims to design a metamaterial absorber with a simple structure, high absorptivity, and broad operational bandwidth, which is expected to have important practical application prospects in infrared imaging, infrared stealth, infrared sensing and other fields.

To address the above issues, a broadband high-absorption embedded multilayer metamaterial absorber is proposed based on the concepts of multimode resonance coupling, composite dielectric layer regulation, and impedance matching theory. First, through the design of multilayer resonant units and embedded cylindrical structures, localized surface plasmon resonance (LSPR), propagating surface plasmon resonance (PSPR) and cavity resonance are simultaneously excited within the subwavelength unit, enabling broadband superposition of multiple resonance peaks. Second, a Ti-Si₃N₄ composite structure is adopted, which introduces a low-permittivity Si₃N₄ layer and a high-loss metal Ti layer. The Ti layer enhances the local confinement of the surface plasmon resonance (SPR) field, while the Si₃N₄ layer broadens the bandwidth through intrinsic absorption and energy coupling effects. Meanwhile, the optimized impedance matching design for the two material layers effectively reduces interface reflection loss. Finally, the centrosymmetric design of the unit structure ensures excellent absorption performance over an incident angle range of 0° to 60°, satisfying the angle adaptability requirements in complex application scenarios. The finite-difference time-domain (FDTD) method is used to analyze the absorption performance and explore the physical mechanisms underlying the broadband high-absorption characteristics, providing a theoretical basis for the design of the absorber.

and Discussion The resonant unit of the metamaterial absorber consists of stacked Si₃N₄, Ti, and Si₃N₄ rectangular layers, with Ti cylindrical structures integrated within the upper Si₃N₄ layer. FDTD simulations show that the absorption spectrum of the absorber exhibits absorption peaks of 99.5%, 99.9%, and 99.8% at wavelengths of 8.61 μm, 14.76 μm, and 15.27 μm, respectively. In addition, the absorber maintains an absorption of over 90% in the wavelength range of 7.92 μm to 16.45 μm, corresponding to an absorption bandwidth of 8.53 μm, and exhibits excellent impedance matching within this band. The absorber features polarization insensitivity, with an average absorption of 91.9% across the 0° to 90° polarization angle range. This polarization insensitivity is mainly due to the structural symmetry of the designed absorber, which results in similar optical responses at all polarization angles, thus reducing the influence of polarization angle. Furthermore, the absorber exhibits wide-angle absorption characteristics; under 60° oblique incidence, the average absorptivity for the TM and TE modes remains at 90% and 76%, respectively. Mechanism analysis demonstrates that the broadband and high-absorption characteristics of the absorber mainly arise from the synergistic coupling of LSPR, cavity resonance, and PSPR.

Based on multilayer stacked structures and impedance matching theory, a broadband high-absorption metamaterial absorber is designed, whose unit consists of a three-layer Si₃N₄-Ti-Si₃N₄ stack with Ti cylinders embedded in the upper Si₃N₄ layer. The high-loss metal Ti can excite SPR to enhance broadband absorption, and Si₃N₄ exhibits favorable intrinsic absorption and SPR coupling effects in the infrared band; thus, the Ti-Si₃N₄ composite structure achieves improved absorption efficiency and a broadened absorption bandwidth. FDTD simulations show that the designed absorber exhibits excellent absorption performance in the 7 μm to 18 μm band, with an absorption rate exceeding 90% in the wavelength range of 7.92 μm to 16.45 μm, corresponding to an absorption bandwidth of 8.53 μm. The adoption of a symmetric patterned structure design endows the absorber with polarization insensitivity, and the average absorption rate remains at 91.9% at all polarization angles. In addition, the absorber also demonstrates excellent wide-angle absorption performance; under 60° oblique incidence, the average absorption rates of the TM and TE modes reach 90% and 76%, respectively. Mechanism investigation demonstrates that the broadband and high-absorption characteristics of the absorber mainly arise from LSPR excited on the surface of Ti cylinders, cavity resonance between units, and PSPR at metal-dielectric interfaces. The coupling between various resonances further enhances the absorption performance and broadens the absorption bandwidth. This absorber integrates the excellent properties of broadband absorption, high absorption rate, polarization insensitivity, and wide-angle absorption, showing potential application value in fields such as energy harvesting and optical detection. It also provides an important reference for the structural design and performance optimization of broadband high-absorption metamaterial absorbers in the infrared band.