The manufacturing of printed circuit boards (PCBs) constitutes the foundational infrastructure of the contemporary electronics industry. As the sector evolves in response to consumer demands for increasingly miniaturized, sophisticated, and high-performance devices, the physical characteristics of PCBs have undergone significant transformation. Modern boards are characterized by high-density interconnects, multi-layered architectures, and microscopic component placement. In this context, the role of quality control has shifted from a supplementary process to a critical determinant of product viability and yield. Traditionally, defect detection has relied on manual inspection or standard automated optical inspection (AOI) systems. However, these conventional methods are increasingly challenged by the escalating complexity of modern PCB designs. Manual inspection is labor-intensive, subjective, and prone to fatigue-induced errors, rendering it unsuitable for mass production. Similarly, traditional AOI systems based on pattern matching or simple image processing often struggle with the subtle variations and complex backgrounds found in high-density PCBs. Consequently, there is a growing necessity for more intelligent and adaptive inspection solutions capable of discerning defects with high precision. Current state-of-the-art detection methods, particularly those leveraging deep learning paradigms, encounter a distinct set of challenges when applied to the specific constraints of industrial manufacturing environments. The first challenge is the significant computational burden associated with high-precision models. Accurately detecting minute defects typically necessitates high-resolution imagery; however, processing such data with deep convolutional neural networks (CNNs) or complex Transformer models consumes substantial computational resources and memory. This computational demand often translates to prohibitive hardware costs for widespread deployment on factory floors. The second challenge is scale variation, which presents a persistent hurdle. PCB defects manifest in widely varying sizes, from tiny pinholes barely visible to the naked eye to larger structural errors like open circuits. A detection model must possess the multi-scale capability to identify these diverse anomalies with uniform accuracy. The third challenge is the imperative for real-time performance. In high-velocity production lines, inspection algorithms must operate at speeds commensurate with manufacturing throughput to avoid creating bottlenecks. Unfortunately, many existing high-accuracy models are computationally too intensive for real-time inference, forcing manufacturers to make an undesirable trade-off between inspection speed and detection precision. These limitations highlight the need for an optimized architecture that balances resource efficiency with robust detection capabilities.

To address these pressing challenges effectively, this study proposes LightRT-DETR, a lightweight real-time detection transformer specifically optimized for PCB defect detection. Building upon the architectural principles of the real-time detection Transformer (RT-DETR), our approach introduces a series of strategic modifications designed to mitigate computational costs significantly while maintaining, and in some aspects enhancing, detection accuracy. The design of LightRT-DETR is characterized by three primary technical innovations: the implementation of the EfficientFastGLU-18 backbone, the integration of a cascaded group attention (CGA) mechanism, and the development of an efficient selective feature pyramid network (ES-FPN). These components work in concert to refine the model's ability to extract relevant features from complex backgrounds without incurring the heavy computational penalties associated with standard transformer architectures. The overarching goal is to create a model that is both agile enough for edge deployment and accurate enough to meet rigorous quality standards. To resolve the inefficiency of traditional backbones like ResNet-18 when deployed on edge computing devices, we propose the integration of the EfficientFastGLU-18 backbone. This novel structure synthesizes partial convolution (PConv) with convolutional gated linear units (GLU). PConv operates by applying convolution to only a subset of input channels while leaving the remainder untouched for the next layer. This strategy effectively reduces memory access and floating-point operations (FLOPs) without compromising the flow of information. By coupling this efficient convolution strategy with GLU, which provides robust non-linear transformation capabilities, the backbone maintains strong feature extraction abilities. This is particularly crucial for identifying defects against the complex, texture-rich backgrounds of PCBs, where differentiating between a valid trace and a defect requires discerning subtle feature variations. Furthermore, to address the computational expense of standard multi-head attention mechanisms, we introduce a cascaded group attention (CGA) mechanism. Instead of processing all feature channels simultaneously, CGA splits feature maps into distinct groups and processes them through a cascaded structure. This approach allows the network to refine features progressively, as the output of one group's attention computation can inform the next, creating a hierarchical refinement process. This mechanism appears particularly beneficial for defects possessing intricate or fine-grained details, such as "mouse bites" or "spurs", which are frequently overlooked by coarser, global attention mechanisms. Finally, we tackle multi-scale defect detection through the efficient selective feature pyramid network (ES-FPN). This redesigned module employs selective feature integration and incorporates RetBlockC3 modules enhanced with manhattan self-attention (MaSA). Unlike standard self-attention, MaSA is specifically designed to capture long-range dependencies in horizontal and vertical directions. Given that PCB layouts consist largely of orthogonal lines, this domain-specific adaptation enables more effective recognition of defects across widely varying scales by aligning the attention mechanism with the geometric priors of the PCB domain.

The efficacy of LightRT-DETR was validated through extensive experiments on a public PCB defect dataset, revealing a significant improvement in both efficiency and accuracy compared to current state-of-the-art models, including the widely utilized YOLO series. In terms of quantitative performance, LightRT-DETR achieved a mean average precision (mAP50) of 98.3% and a stricter metric, mAP50-95, of 52.9%. When benchmarked against the advanced YOLOv9e model, our approach demonstrated an improvement of 3.0% in mAP50 and 1.8% in mAP50-95. These metrics suggest that the model is not only capable of detecting defects but also does so with high localization accuracy, minimizing false positives and false negatives—a critical requirement for industrial quality control. The experimental setup ensured fair comparisons by utilizing consistent hyperparameters and training environments, further reinforcing the validity of the observed performance gains.

Perhaps the most distinct advantage of LightRT-DETR lies in its computational efficiency. The model requires approximately 37.2 GFLOPs, representing an 80.3% reduction in computational cost compared to the baseline RT-DETR model. Similarly, the parameter count is reduced to 12.8 million, a 77.7% decrease. This substantial reduction has practical implications: it facilitates deployment on cost-effective hardware with limited computing power or allows the system to operate at significantly higher frame rates on standard hardware.

UAV aerial images have been widely applied in military surveillance, emergency rescue, agricultural monitoring, and other fields due to their advantages of flexible shooting angles and rich spatial information acquisition. However, target detection in such images faces prominent challenges: extreme scale variations of targets, dense spatial distribution of small targets, and frequent false or missed detections caused by complex backgrounds. Additionally, real-time application scenarios impose strict requirements on the inference speed of detection models, which makes it difficult for existing algorithms to balance detection accuracy and computational efficiency. To address these issues, this study aims to propose a high-precision and efficient UAV aerial image target detection method, which enhances the model's ability to perceive multi-scale targets, especially small targets, reduces false and missed detections, and ensures real-time performance to meet practical application needs.

Based on the YOLOv11s baseline model, this study designs a multi-module collaborative optimization framework, and the specific technical methods are as follows: First, an Adaptive Multi-dimensional Feature Sampling (AMFS) unit is embedded in the backbone network. The input feature maps are grouped along the channel dimension, and three parallel paths are used for multi-scale feature extraction. Adaptive average pooling, depthwise convolution, and spatial attention mechanisms are integrated to realize the adaptive synergy fusion of local details and global contextual information, which makes up for the limitation of fixed convolution receptive fields and enhances the feature extraction ability for small targets. Second, a Spatial Pyramid Adaptive Pooling Fusion (SPAPF) layer is constructed to replace the traditional SPPF module. This layer combines smoothed approximate average pooling and smoothed approximate maximum pooling through learnable weight masks. It not only retains structural details and texture information of targets but also reduces redundant computations, thus accelerating the model's inference speed while maintaining lightweight characteristics. Third, a Cross-Spatial Contextual Parallel Self-Attention (CS_C2PSA) module is proposed. It introduces cross-spatial multi-head self-attention (CSL-MHSA) and phased residual connection structures. By associating local and global receptive fields through 1×1 and 3×3 convolutions, and integrating vertical and horizontal spatial position information, the module strengthens the model's global contextual perception ability, alleviates the confusion between small targets and backgrounds or adjacent targets, and improves the false detection problem. Finally, a Semantic-guided Feature Fusion (SGFF) module and a dedicated Small Object Detection (SOD_Detect) head are designed. The SGFF module adopts a cross-layer feature alignment strategy and sub-pixel convolution technology. It uses the SE channel attention mechanism to enhance the expression of deep semantic information in shallow features, realizes the efficient fusion of shallow detail features and deep semantic features, and generates high-resolution feature maps with rich semantics for the SOD_Detect head, which significantly improves the positioning and classification accuracy of small targets. The model is trained and tested on VisDrone2019 and VisDrone2021 datasets, with batch_size set to 16 and epochs to 200. The experimental environment is Windows 10 operating system, NVIDIA GeForce RTX 3090 GPU, PyTorch 1.10.1 + CUDA 11.1 deep learning framework, and Python 3.8 programming language.

Ablation experiments on the VisDrone2019 dataset show that each improved module plays a positive role in performance enhancement: The AMFS unit increases the mAP@0.5 by 3.2%, the SPAPF layer reduces the GFLOPs by 4.4 while improving the Recall by 2.2%, the CS_C2PSA module increases the Precision by 2.3% and the mAP@0.5 by 2.0%, and the SGFF module improves the Precision by 3.9% and the mAP@0.5 by 3.4%. When all four modules are integrated, the model's Precision reaches 58.3%, Recall 43.1%, and mAP@0.5 46.8%, which are 8.1%, 4.7%, and 7.2% higher than those of the baseline YOLOv11s, respectively, and the GFLOPs are maintained at 30.8, ensuring computational efficiency. Comparative experiments with 11 mainstream algorithms (including SSD, Faster-RCNN, YOLOv5s, YOLOv11s, etc.) on VisDrone2019 and VisDrone2021 datasets show that the proposed method achieves mAP@0.5 of 46.8% and 44.7% on the two datasets, respectively, which are 7.2% and 6.5% higher than the baseline model. Compared with the latest YOLOv10s and YOLOv8s, the mAP@0.5 is increased by 8.0% and 8.6% respectively on VisDrone2019 dataset. Compared with the UAV-specific detection algorithm ESO-DETR, the mAP@0.5 is improved by 2.1%, and the FPS is increased by 14. Even compared with the lightweight model PP-YOLOE, the proposed method still has advantages in both accuracy and speed, with Precision increased by 4.3% and FPS increased by 21. Visualization results of challenging scenarios such as dense small targets, aerial multi-scale, and low-light occlusion show that the proposed method effectively reduces false and missed detections, and has better detection performance for dense small targets than algorithms such as MIS-YOLOv8. However, there is still room for improvement in the distinction between Pedestrian and People categories in extremely dense scenarios, which is a direction for subsequent optimization.

This study proposes a UAV aerial image target detection method based on semantic guidance and multi-dimensional feature perception. Through the collaborative optimization of AMFS, SPAPF, CS_C2PSA, and SGFF modules, the method effectively solves the key problems of large target scale variation, dense small target distribution, and high false and missed detection rates in complex backgrounds in UAV aerial image target detection. Experimental results on public datasets verify that the proposed method significantly improves the detection accuracy while maintaining efficient computational performance, and has excellent generalization ability and robustness. It provides a reliable technical solution for high-precision and real-time target detection in UAV aerial imagery. In future research, we will further optimize the module structure to enhance the model's ability to distinguish similar target categories in extremely dense scenarios, and explore the application of the method in more complex environments such as severe weather and heavy occlusion to expand its practical application scope.

Small object detection remains a significant and persistent challenge within the computer vision domain. This challenge is particularly acute in the context of drone aerial imagery, where objects captured from high altitudes appear extremely small, densely packed, and often exhibit blurred features due to low resolution and environmental factors. Traditional object detection algorithms frequently suffer from performance degradation in these scenarios, manifesting as high rates of missed detections and false positives. To address these critical limitations, this paper proposes a novel object detection algorithm named FFC-YOLO (FASFF-FreqFusion-CAA YOLO), which is built upon a frequency-aware feature fusion framework.

The core of the FFC-YOLO algorithm lies in three systematic and synergistic innovations designed to enhance feature representation, fusion, and contextual understanding specifically for small and dense objects. First, to bolster multi-scale feature extraction, we redesign the detection head by incorporating a Four Adaptively Spatial Feature Fusion (Detect-FASFF) structure. This is augmented with an additional P2 detection layer dedicated to capturing finer-grained spatial information crucial for identifying minuscule targets. The FASFF mechanism leverages four adaptive branches to dynamically aggregate features across different scales, thereby significantly improving the model's sensitivity and discriminative power for small objects across varying sizes.Second, we comprehensively optimize the feature fusion pathway to counteract the information loss common in conventional sampling methods (e.g., up-sampling and down-sampling). We introduce a Frequency-aware Feature Fusion (FreqFusion) module. This module operates by decomposing features into different frequency components within the frequency domain, allowing for more deliberate and effective integration of high-frequency details (essential for edge and texture information of small objects) and low-frequency semantics. This FreqFusion module is then elegantly coupled with a Bidirectional Feature Pyramid Network (BiFPN) structure. The resulting fusion architecture enables efficient, multi-level bidirectional cross-scale information flow, effectively mitigating the problems of feature obliteration, category misclassification, and bounding box localization drift in scenes crowded with small objects.

Third, to enhance the model's ability to leverage surrounding semantic context—a vital clue when object appearance is ambiguous—we integrate a Context Anchor Attention (CAA) mechanism into the C2PSA module. This attention mechanism functions by establishing dynamic associations between anchor points and their contextual surroundings, enabling the feature extraction process to selectively focus on and amplify informative regions around potential targets. This contextual awareness strengthens feature representation and aids in distinguishing small objects from complex backgrounds.

The efficacy of the proposed FFC-YOLO algorithm is rigorously validated through extensive experiments on two benchmark datasets. On the public VisDrone2019 dataset, a standard benchmark for drone vision, FFC-YOLO achieves a mean Average Precision (mAP@0.5) of 40.0%. This represents a substantial performance gain of 8.0%, 7.6%, and 7.8% over the widely-recognized YOLOv8n, YOLOv10n, and YOLOv11n baselines, respectively. To further verify its generalization capability and prowess in detecting extremely small targets, we conducted additional experiments on a custom-built, challenging dataset named "tiny-data", which contains three categories of progressively smaller persons ("sperson", "lperson", "wperson"). On this dataset, FFC-YOLO demonstrates superior results compared to the strong YOLOv11n baseline, with improvements of 9.2% in mAP@0.5, 8.7% in Precision (P), and 5.9% in Recall (R). These consistent gains across both public and private datasets robustly confirm that the FFC-YOLO algorithm possesses strong generalization ability and excels in the demanding task of drone-based aerial image analysis, particularly for long-range and dense small object detection.

In conclusion, this paper presents a comprehensive and effective solution to the small object detection problem in aerial imagery. By innovating at the levels of feature extraction (FASFF head), feature fusion (FreqFusion-BiFPN), and contextual modeling (CAA attention), the FFC-YOLO framework successfully addresses key shortcomings of prior methods. The experimental evidence underscores its potential for practical application in UAV-based surveillance, traffic monitoring, and search-and-rescue operations. For future work, we plan to explore model compression techniques such as pruning and knowledge distillation to streamline the model for efficient deployment on resource-constrained edge devices, thereby bridging the gap between high accuracy and real-time performance.

In complex real scenes (such as low illumination, haze, occlusion or strong light interference), single mode target recognition and localization methods face significant performance bottlenecks. Although the visible image is rich in texture and color details, it is highly sensitive to light conditions; Infrared image has all-weather working ability and strong penetrability by virtue of thermal radiation imaging, but it is limited by low resolution and lack of structural information. The inherent limitations of the two make it difficult for the single-mode system to meet the requirements of all-weather perception with high robustness and high accuracy. Therefore, how to deeply integrate the complementary advantages of infrared and visible modes, that is, combining the local detail expression ability of visible light with the global thermal semantic saliency of infrared, has become a key challenge to improve the performance of target detection in complex environments. This paper aims to build a new neural network architecture that can deeply mine, adaptively fuse and efficiently use multimodal features, in order to achieve higher accuracy, stronger robustness and better scale invariance of target recognition and localization.

This paper proposes an infrared visible multimodal target recognition and positioning network based on CNN transformer heterogeneous fusion framework. The network is composed of four core modules: visible light oriented CNN feature extraction branch, infrared oriented transformer feature extraction branch, cross modal adaptive fusion module and multi-scale dynamic adjustment structure. In the visible light branching, the "feature criticality grouping convolution" mechanism is designed: the feature importance is evaluated by weighting the maximum and average values at the channel level, and the normalized criticality weight is generated after full connection layer optimization. Based on this, the number of convolution output channels is dynamically allocated - the key region gets richer characterization, and the background region is compressed, so as to guide the network to focus on the essential information of the target; At the same time, the compression extraction module composed of 1 × 1 point convolution and Gelu activation is introduced, and the hybrid down sampling strategy is adopted to reduce the computational overhead and the loss of micro target information. In the infrared branch, the feature aggregation shift window self-attention structure is constructed: the computational complexity is controlled by using the window shifting mechanism of swing transformer for reference, and 2 × 2 average pooling is applied to the key (k) and value (V) matrices in the self-attention for local aggregation, which not only significantly reduces the redundant calculation, but also strengthens the ability of global context modeling for low contrast heat source targets. Then, a cross modal fusion module is designed: firstly, the local features of CNN and the global features of transformer at the same scale are normalized, and then channel attention (highlighting semantic importance) and spatial attention (accurately locating the target area) are applied respectively, and shallow features are fused to enhance detail retention; Finally, the adaptive stitching of the three features is realized through the learnable weight parameters, and the complementary enhancement between modes is achieved. Finally, in order to alleviate the feature conflict in multi-scale fusion, a multi-scale dynamic adjustment mechanism is proposed: Based on the feature pyramid structure, the fusion proportion is dynamically adjusted according to the difference in the number of adjacent scale channels (the larger the scale difference, the smaller the fusion weight), and the channel spatial attention mechanism is combined to strengthen the target response, suppress the background interference, and significantly improve the unified detection ability of small, medium and large targets.

Experiments is systematically validated on three public datasets—KAIST, FLIR, RGB-T—and a self-collected power grid equipment dataset (covering 5 types of electrical equipment). The results show that: 1) the proposed key-grouped convolution achieves 83.1% mAP on the visible light modality, with parameters and FLOPs reduced by 26% and 21% compared to standard convolutions, and inference speed increased by 29%, outperforming mainstream lightweight networks and demonstrating an excellent efficiency-accuracy balance. 2) After introducing feature aggregation in the infrared branch, FLOPs decreased from 173.2 G to 116.5 G, FPS increased from 69 to 88, and mAP only slightly dropped by 0.9%, verifying its efficiency. 3) The modality-network adaptability experiment clearly shows that CNN performs better on visible light (83.1% mAP) than Transformer (81.2%), while Transformer outperforms CNN on infrared (78.9% vs. 78.2%), strongly supporting the rationality of the heterogeneous dual-branch design. 4) In the comparison of multimodal fusion strategies, the proposed attention-guided adaptive fusion achieved 88.4% mAP on KAIST, significantly surpassing addition (86.8%) and concatenation (87.5%) methods; feature visualization confirms its effectiveness in suppressing background noise, particularly benefiting small object edge preservation. 5) The multi-scale fusion module, built on FPN and ASFF, introduces dynamic ratio adjustment and dual attention, achieving an overall mAP of 89.3%, with small object mAPs improved to 79.4%, validating its effectiveness in alleviating scale conflicts. 6) Comprehensive comparisons with five advanced methods (including dual CNNs, dual Transformers, CNN-Transformer cascades, etc.) across the four datasets show that our method achieved the best or second-best results on KAIST (89.3%), FLIR (88.9%), RGB-T (87.1%), and the power grid dataset (89.1%), with controlled computational cost (FLOPs ≈169 G, FPS ≈69), balancing accuracy, robustness, and practicality. In summary, this method systematically addresses the issues of insufficient multimodal feature complementarity and scale sensitivity through the deep synergy of modality characteristics, network architecture, and fusion strategies.

Aiming at the problems of insufficient feature mining, weak modal complementarity and poor multi-scale adaptability in infrared visible multimodal target recognition and localization, this paper innovatively proposes a heterogeneous network architecture that deeply integrates the advantages of CNN and transformer. Through visible light oriented feature criticality grouping CNN, infrared oriented feature aggregation lightweight transformer, attention driven adaptive cross modal fusion and dynamic multi-scale adjustment mechanism, the extraction efficiency, fusion quality and scale invariance of multimodal features are systematically improved. A large number of experiments have verified the superior performance of this method on multiple real scene data sets, which not only leads significantly in detection accuracy, but also maintains good computational efficiency and environmental robustness. This study provides an effective technical path for highly reliable target perception in complex environments. Future work will focus on model compression (such as knowledge distillation, neural architecture search) to adapt edge devices, and explore strategies to enhance the generalization ability in extreme weather (such as rainstorm and smoke), so as to further promote the application of this technology in key fields such as power patrol, automatic driving, emergency rescue and so on.

Residential gas leaks, vehicle exhaust, and industrial emissions all contain carbon monoxide (CO), a colorless, odorless, and extremely poisonous gas produced by the incomplete combustion of fuels containing carbon. Public safety, industrial production, and environmental quality are all seriously threatened by CO, which can cause acute poisoning or even death in humans at very low amounts. Therefore, it is crucial to create ultra-sensitive CO detection technology to protect human health, industrial operations, and environmental monitoring. Researchers have recently focused a lot of attention on anti-resonant hollow-core fiber (AR-HCF) because of its special benefits, which include low transmission loss, large core diameter, small optical stripe amplitude, compact structure, and the ability to enable long-distance continuous interaction between single-mode radiation and the target gas.These features give AR-HCF a lot of potential for use in gas sensing applications. In order to achieve high sensitivity, high stability, and compact trace CO detection, this study proposes an all-fiber gas detection system based on AR-HCF for CO monitoring employing fiber loop ring down spectroscopy.

In this work, an all-fiber CO detection system based on fiber ring cavity decay spectroscopy was built, using a continuous wave (CW) laser with a central wavelength of 1567 nm as the light source and AR-HCF as the gas chamber sensing element. The AR-HCF' transmission characteristics were first examined using the finite element approach, with an emphasis on its limiting loss at the operating wavelength of 1567 nm.The ideal length of the AR-HCF gas chamber was then determined by methodically examining the link between the sensitivity of the hollow-core gas chamber and the optical path length. Erbium-doped fiber amplifier (EDFA) with varying gain values were incorporated into the fiber ring cavity to boost the effective number of laser cycles and offset the intrinsic cavity loss in order to further improve the detection sensitivity. In order to thoroughly assess the detection performance of the developed system, a number of performance tests were carried out, including stability testing through repeated measurements and concentration detection tests in the range of 0-0.02%.

The research using the finite element approach showed that the AR-HCF' limiting loss at 1567 nm was roughly 0.92 dB/km.A 1-meter-long AR-HCF was chosen as the gas chamber structure after research into the relationship between optical path length and sensitivity revealed that the detection sensitivity increased with the length of the optical path. This was done while taking system compactness and practical applicability into full consideration. The findings of the experiment showed that increasing the EDFA gain greatly increased the system's sensitivity: the system's sensitivity with a 2 dB gain ring cavity was 16.7 times greater than that of the system without EDFA. With a correlation value (R²) of 0.99989, concentration detection experiments verified a strong linear relationship between the decay time and CO concentration in the range of 0%-0.02%. The greatest average relative error was only 0.4%, according to stability tests with ten repeated measurements for various CO concentration gradients. In the end, the developed system demonstrated outstanding detection performance and reliability with a minimum detection limit of 3.0757×10⁻¹¹, a sensitivity of 0.795 s, and a stability of 0.82%.

This work effectively creates an all-fiber CO detection system using fiber loop ring down spectroscopy in conjunction with AR-HCF. The system achieves ultra-high sensitivity and stability in trace CO detection by utilizing the superior performance of AR-HCF and the gain compensating effect of EDFA. The system offers a new and efficient technical method for trace gas detection thanks to its benefits of strong anti-interference ability, compact structure, high detection sensitivity, and dependable performance.It not only improves the technical methods of CO detection but also has great potential for use in areas including public health protection, environmental quality monitoring, and industrial safety early warning. In order to increase the detection limit and broaden the system's application scope to more trace gas detection scenarios, future research can concentrate on improving the fiber cavity construction and EDFA gain control.

Real-time and accurate detection of rockfalls on mining roads is critical for ensuring transportation safety, preventing geological hazards, and maintaining operational continuity in open-pit mining environments. Detecting rockfalls in such scenarios presents substantial challenges due to the small size of targets, irregular shapes, low contrast against the background, drastic illumination variations, and interference from adverse weather conditions such as fog, dust, and rain. Conventional single-modality visual detection methods, particularly those based solely on visible light imagery, often fail under low-visibility conditions, while existing dual-modality fusion approaches typically face limitations including complex network structures, high computational costs, and difficulties in simultaneously achieving real-time performance and high detection accuracy for small targets.

To address these challenges, this study proposes a novel progressive infrared–visible light fusion detection framework based on the Mamba state space model (SSM), which is specifically designed to efficiently integrate complementary information from infrared and visible light images while preserving fine-grained details of small targets, even under challenging environmental conditions such as low illumination, strong backlighting, partial occlusion, and adverse weather. The framework consists of three key components: 1) optimized feature extraction using MobileMamba and Wavelet Transform-based downsampling, which reduces feature loss while maintaining high-resolution details, 2) cross-modal feature fusion via the MambaFusion module, enabling adaptive integration of thermal and texture cues across multiple scales, and 3) small-object detection and localization enhancement through a Scale and Location Sensitive Loss (SLS Loss), which explicitly models scale variations and center-point offsets to improve both classification confidence and spatial precision for micro-scale rockfall targets. This multi-component design ensures robust detection in complex, unstructured mining environments and provides a balanced trade-off between computational efficiency and detection accuracy.

First, a Wavelet Transform-based WTSterm downsampling module is introduced to mitigate the loss of fine-grained features during feature extraction. Unlike conventional stride-based convolutional downsampling, the WTSterm module decomposes input images into multiple frequency components, including low-frequency approximations and directional high-frequency details. These components are concatenated along the channel dimension and projected through a 1×1 convolution, effectively preserving edge and texture information of small rock targets while reducing computational overhead. This module is integrated into an optimized MobileMamba backbone network, a lightweight visual SSM architecture that combines multi-kernel convolutions, multi-receptive field feature interaction, and wavelet-enhanced Mamba blocks. The MobileMamba backbone enables efficient extraction of thermal radiation cues from infrared images and rich spatial-texture features from visible light images, while selectively maintaining both shallow fine-grained details and deep high-level semantic features. To reduce computational redundancy, the network removes high-dimensional, small-scale feature branches that contribute little to small-object representation, thereby balancing feature richness and efficiency.

Second, a MambaFusion module is designed to progressively fuse multi-scale features from infrared and visible light modalities. The fusion process leverages Mamba’s dynamic state transition and gating mechanisms, which enable adaptive selection of discriminative information while suppressing modality-specific noise. MambaFusion combines the global dependency modeling ability of the SSM with local feature retention through convolution, facilitating cross-modal correspondence at semantic and spatial levels. Multi-scale fusion is performed across large, medium, and small feature maps, with adaptive weighting applied to emphasize salient information according to scene context. This mechanism allows the network to maintain robustness under challenging conditions such as strong backlighting, low illumination, partial occlusion, and airborne dust or smoke, enhancing both target discrimination and representation for small rockfall instances.

Third, to improve small-object detection and localization, the framework incorporates a Scale and Location Sensitive Loss (SLS Loss). This loss function explicitly models variations in target scale and center-point offsets, addressing the insensitivity of conventional losses (e.g., IoU, Dice) to size and position discrepancies. The scale-sensitive component reweights loss contributions according to the predicted and true target sizes, focusing training on underrepresented or mispredicted small targets. The location-sensitive component employs polar-coordinate quantification of center-point offsets, providing a strong penalty for positional errors and refining bounding box localization. Together, SLS Loss ensures that the network maintains high sensitivity to micro-scale rockfalls and achieves precise positioning in cluttered or unstructured road environments.

Extensive experiments were conducted on a self-collected open-pit mine rockfall dataset, containing diverse scenes with varying illumination, weather conditions, and target sizes (typically 10–50 cm in diameter). Results demonstrate that the proposed framework achieves precision of 0.856, recall of 0.715, and mAP@0.5 of 0.721, with an inference speed of 16 ms per frame. Compared with conventional single-modality detection and existing dual-modality fusion approaches, the proposed method significantly improves small-target detection performance and robustness while maintaining real-time computational efficiency. In conclusion, the progressive infrared–visible light fusion framework based on MobileMamba and MambaFusion modules effectively integrates complementary thermal and textural cues, preserves fine-grained details of micro-scale rockfall targets, and enhances small-object localization and discrimination in complex visual scenarios. By combining efficient feature extraction, adaptive cross-modal fusion, and a scale- and location-aware loss function, the proposed approach achieves high robustness under diverse environmental conditions, including low visibility, dynamic lighting, dust, and partial occlusion.

The method offers a reliable, high-performance solution for real-time rockfall detection and monitoring in open-pit mining environments, significantly contributing to safer road operation, reduced operational risk, and improved transportation efficiency. These results demonstrate the potential of lightweight state-space models combined with progressive multi-modal fusion and targeted loss functions for deployment in practical industrial safety and autonomous monitoring systems.

Photonics integrated circuits (PICs) have emerged as a foundational platform for next-generation optical information transmission and processing, owing to their high speed, broad bandwidth, excellent scalability, and low propagation loss. These advantages stem from the unique capability of PICs to densely integrate diverse and complex optical functionalities onto micro- and nanoscale chips. However, the efficient conversion of guided optical modes into free-space radiation propagating into the far field, together with precise and flexible manipulation of multiple optical degrees of freedom (DoFs)—including amplitude, phase, state of polarization (SoP), and orbital angular momentum (OAM)—remains a critical challenge in the field. Addressing this challenge is essential for unlocking the full potential of integrated photonic systems. In this context, the present review provides significant academic and technological value by systematically summarizing recent advances in on-chip metasurfaces for guided-mode radiation control. It fills an important gap in the existing literature by consolidating both theoretical frameworks and experimental demonstrations, while offering a comprehensive roadmap for overcoming key technical bottlenecks. By doing so, this review accelerates the translation of innovative metasurface concepts into practical PIC applications and supports the development of intelligent, high-performance optical systems.

Over the past decade, integration strategies for metasurface–waveguide architectures have evolved substantially. Early implementations typically involved placing a metasurface layer atop a conventional grating coupler. As demands for miniaturization and functional density increased, more advanced approaches emerged, most notably the direct patterning of high-aspect-ratio, high-refractive-index artificial nanostructures (meta-atoms) on waveguide surfaces. This strategy enables direct modulation of the evanescent field of guided modes, thereby facilitating comprehensive control over multiple DoFs of the radiated optical field. Recent progress in PICs and nanophotonics has further inspired innovative design paradigms, among which two representative approaches have attracted particular attention: co-designed metasurface–grating coupler hybrids and leaky-wave metasurfaces. The former offers a compact and fabrication-friendly solution, often realizable via a single etching step, in which the metasurface–grating coupler simultaneously enables efficient radiation and wavefront shaping. The latter exploits intentional symmetry breaking in photonic crystal slab waveguides to induce controlled leaky-wave radiation. Notably, both strategies can realize large surface-emitting apertures without requiring the integration of high-aspect-ratio, high-index nanostructures, making them especially attractive for scalable implementations. As a result, these two approaches are expected to play a central role in advancing the field.

Substantial progress has been achieved in manipulating guided-mode radiation using on-chip metasurfaces. Several well-established mechanisms have been proposed to control the phase of radiated wavefronts, including: 1) detour phase, determined by the lateral displacement of a meta-atom within a unit cell; 2) geometric (Pancharatnam–Berry) phase, governed by the in-plane rotation of anisotropic meta-atoms; 3) propagation phase, set by the geometric dimensions of the meta-atoms; and 4) resonant phase, arising from engineered resonances that provide strong and dispersive phase responses. Beyond phase-only modulation, interference among multiple artificial nanostructures within a unit cell enables simultaneous control of amplitude and SoP, thereby allowing full multidimensional manipulation of optical DoFs. By leveraging these mechanisms, on-chip metasurfaces have demonstrated a broad range of functionalities, including beam steering, focusing, vortex beam generation, and holographic projection. Furthermore, the inherent advantages of the on-chip platform—such as multi-port, multi-mode, and multi-wavelength excitation—enable decoupled and independent control of optical DoFs with minimal crosstalk, paving the way for advanced functional multiplexing in integrated photonic systems.

Despite these advances, several key challenges remain. First, the radiation efficiency of guided modes must be further improved. Future studies are expected to explore novel radiation-enhancement mechanisms, such as resonant architectures and multilayer composite structures, to increase energy conversion efficiency. Second, modulation precision is still constrained by current micro- and nanofabrication limitations. Advances in stacked multilayer metasurfaces, three-dimensional (3D) nanostructures, and high-precision, multi-material co-fabrication could significantly expand the available design space and improve device fidelity. Third, dynamic modulation capabilities remain limited. Most existing on-chip metasurfaces are passive and rely on static geometries, preventing adaptive responses to external stimuli such as temperature, optical intensity, or applied voltage. Future efforts may enable non-volatile reconfigurability through phase-change materials, achieve ultrafast modulation using the optoelectronic properties of two-dimensional (2D) materials, or expand modulation dimensionality via photochromic effects. Finally, current on-chip metasurfaces often lack sufficient system-level integration. Future research should therefore focus on the co-integration of metasurfaces with multi-functional photonic modules to realize intelligent, system-level optical field manipulation.

In conclusion, pioneering work on on-chip metasurfaces has significantly advanced the understanding of guided-mode radiation control and driven innovation in PICs and nanophotonics. With continued progress in nanofabrication technologies and electromagnetic modeling, on-chip metasurfaces are expected to achieve further performance improvements. As a result, metasurface-enabled control of guided-mode radiation is poised to become a cornerstone technology for next-generation PICs, with far-reaching impact across high-capacity optical communications, imaging systems, optical security, and immersive AR/VR applications.

As a core technology in the field of information security, image steganography holds irreplaceable strategic value in areas such as covert communication and digital copyright protection. Current adversarial image steganography methods commonly suffer from insufficient cross-model generalization capability and low quality of stego images. This is primarily due to the over-reliance of these methods on the decision boundaries of the target steganalysis model, which results in adversarial perturbations being highly coupled with the decision boundaries of the target model, making it difficult to generalize effectively to unknown models. Moreover, when generating adversarial perturbations, the fusion of secret information with the texture of the cover image is often insufficiently natural, especially in low-texture regions where statistical anomalies are prone to emerge, thereby compromising the visual quality and statistical naturalness of the stego images.

To address the aforementioned issues, this paper proposes a dual-strategy image steganography model that integrates passive defense and active attack. Through the synergistic end-to-end optimization training of cover image enhancement and adversarial embedding, the model significantly improves cross-model generalization capability and the quality of stego images. At the passive defense level, cover image optimization is performed: A dual-stream U-Net structure is introduced into the generator of the generative adversarial network (GAN). This dual-branch network simultaneously processes the original image and edge information derived from conformal monogenic phase congruency (CMPC). Combined with the SE-Net attention mechanism to optimize feature distribution, texture enhancement and detail preservation are achieved during adversarial training. By jointly optimizing the discriminator loss and the pixel-space mean squared error loss, an optimized cover image with high visual quality and statistical naturalness is generated, constructing a secure and robust cover environment for secret information embedding. At the active attack level, adversarial embedding is conducted: integrated gradients are utilized for neuron attribution to locate the critical discriminative features commonly relied upon by various models, followed by dynamic perturbation. The neuron attribution loss, together with the discriminator loss and pixel-space mean squared error loss, jointly optimizes the embedding strategy. This ensures the indistinguishability between the stego image and the original cover image while significantly reducing the detection accuracy of various steganalysis models. Considering the phased differences in the optimization objectives of the proposed model, the cover image optimization network dominates the training in the initial stage, as a high-quality cover image is the foundation of steganography. In the later stage, the adversarial embedding network takes the lead, continuously adjusting the embedding distortion by interfering with the discriminative logic of steganalysis models to enhance the concealment of the stego image. Therefore, a linear dynamic adjustment strategy is adopted, allowing the weights of the two components to transition smoothly and uniformly during the first 100 training epochs. This matches the requirement for phased dominance switching between the two networks, driving the model's performance to advance progressively by stages and achieving a global optimization of concealment, usability, and security.

In the loss function weight configuration experiment, considering the phased differences in the model's optimization objectives, a linear dynamic adjustment strategy was adopted to change the weight coefficients k₁ and k₂, while systematically investigating the effects of different configurations of λ₁ and the ratio λ₂:λ₃:λ₄. By comparing the quality of stego images and the detection accuracy of non-targeted steganalysis models, the optimal weight configuration was determined as: k₁ decreasing from 0.8 to 0.2, k₂ increasing from 0.2 to 0.8, λ₁=0.2, and λ₂:λ₃:λ₄=8:3:1. The ablation study, by comparing the ACC against non-targeted steganalysis models as well as the average PSNR and SSIM of stego-images under different architectures, demonstrates that the configuration of the proposed method achieves the best overall performance. Furthermore, comparing the effects of adding the SE-Net attention mechanism at different positions within the dual-stream U-Net network showed that adding it only during the encoding stage yielded the best results. In the experiment analyzing generated image quality and efficiency, the average PSNR and SSIM obtained from 1000 randomly selected stego images were 40.89 dB and 0.9783, respectively. Three sets of cover images and their corresponding stego images were randomly selected for comparison. The results indicated that the two types of images are not only visually indistinguishable but also exhibit highly similar histogram distributions. To compare generation efficiency, the average generation time for a single stego image was tested. Due to its complex network architecture, the proposed method is slightly less efficient in generation speed compared to other schemes, but it significantly enhances the quality of cover images. This demonstrates that the proposed method provides an effective guarantee for improving security and image quality at an acceptable efficiency cost. The cross-dataset generalization capability analysis experiment showed that the proposed method achieved the lowest average ACC, indicating good cross-dataset generalization ability. This is because the proposed dual-strategy model generates optimized covers with consistent structural properties and perturbs common discriminative features, effectively reducing the model's dependence on statistical features specific to a particular training dataset. Consequently, it exhibits stable and good anti-detection performance across different datasets. In the experiment on anti-steganalysis capability, a comparison was made with two state-of-the-art adversarial embedding methods (CR-AIS and Natias). The experimental results showed that, in terms of resilience against non-targeted steganalysis models, the proposed method achieved average detection accuracy (ACC) reduction margins of 3.69% and 1.91% compared to CR-AIS and Natias, respectively. This indicates that the proposed method possesses strong resilience against non-targeted steganalysis models and exhibits good cross-model generalization capability.

Through the deep integration of passive defense (dual-stream U-Net-guided cover optimization) and active attack (neuron attribution-driven adversarial embedding), the proposed image steganography model significantly reduces the detection accuracy of various steganalysis models, enhances cross-model generalization capability, and maintains high image quality, thereby achieving synergistic improvement in steganographic security and visual fidelity.

Forged parts are widely used as key load-bearing components in automotive, aerospace, shipbuilding, and heavy equipment industries. Their surface quality directly affects structural reliability and service safety. During forging and subsequent processing, surface defects such as cracks, inclusions, pores, oxide layers, and spot-like imperfections are frequently generated due to material heterogeneity, mold wear, and process instability. Magnetic particle inspection is commonly adopted for ferromagnetic forged parts because of its high sensitivity to surface and near-surface cracks. However, magnetic particle inspection images usually exhibit complex backgrounds, uneven illumination, magnetic trace interference, and weak texture contrast. These characteristics cause surface defects to present blurred boundaries, irregular shapes, and large scale variation. Slender cracks and tiny defects are especially difficult to detect accurately under real-time constraints. Existing convolutional neural network–based and Transformer-based detection models often show insufficient directional perception, limited contextual representation, and inadequate geometric sensitivity in bounding-box regression. As a result, detection accuracy and computational efficiency are difficult to balance in practical industrial inspection scenarios. The objective of this study was to develop a real-time forged-part surface defect detection model that enhances directional feature perception and contextual fusion capability, improves multi-scale feature alignment, and strengthens localization robustness for elongated defects, while maintaining high inference efficiency suitable for industrial deployment.

A real-time surface defect detection framework, named MCF-RTDETR, was constructed based on an improved RT-DETR architecture. Magnetic particle inspection images of heavy-duty truck steering knuckles were collected from an industrial production environment and manually annotated to establish a forged-part surface crack dataset. The dataset contained complex background structures and fine-scale defect patterns, providing realistic samples for model evaluation.

In the backbone network, a multi-gradient fusion convolution module was introduced to enhance directional edge representation. This module integrated four directional difference convolutions with one standard convolution to capture gradient responses along multiple orientations. Such design improved feature sensitivity to slender cracks and weakly textured defects. To avoid increased inference cost, convolutional re-parameterization was adopted during the inference stage, allowing the multi-branch structure to be equivalently transformed into a single convolution layer while preserving feature enhancement capability.

In the neck network, a context-aware recursive feature pyramid network was designed to address insufficient spatial–semantic alignment in conventional multi-scale fusion schemes. A rectangular self-calibration module was first employed to model directional contextual information by expanding receptive fields along horizontal and vertical directions, which strengthened anisotropic defect representation. Subsequently, a FuseBlockMulti module was used to realize semantic-guided alignment between shallow detail features and deep semantic features, enabling adaptive information interaction across different feature levels. Furthermore, a dynamic interpolation fusion mechanism was introduced to assign adaptive fusion weights to multi-scale features according to local contextual semantic strength. Through recursive fusion, feature consistency across scales was progressively refined, improving robustness under complex backgrounds and large scale variation.

For bounding-box regression, the original loss function was replaced with a Focaler-MPDIoU loss. This loss combined difficulty-aware sample weighting with multi-point distance geometric constraints, enhancing localization accuracy and boundary fitting ability for elongated and irregular defects. The entire detection framework was optimized in an end-to-end manner and evaluated under unified training and inference configurations.

Experimental results demonstrated that the proposed MCF-RTDETR achieved superior performance in forged-part surface defect detection. On the magnetic particle inspection dataset of heavy-duty truck steering knuckles, the proposed model obtained a mean average precision of 88.5%, which exceeded the RTDETR-r18 baseline by 2.3%. Meanwhile, computational complexity was reduced from 58.3 G to 33.5 G, and inference speed reached 137.2 f/s, indicating an effective balance between detection accuracy and computational efficiency.

Performance improvement was particularly evident for slender cracks and small-scale defects. Enhanced directional gradient perception reduced missed detections caused by weak edge responses, while context-aware recursive fusion suppressed background interference and improved defect-background discrimination. Ablation experiments confirmed the effectiveness of each proposed component. The multi-gradient fusion Convolution strengthened edge sensitivity without increasing inference overhead. The Context-Aware Recursive Feature Pyramid Network improved multi-scale feature consistency and spatial-semantic alignment. The Focaler-MPDIoU loss further enhanced bounding-box regression accuracy for defects with large aspect ratios.

Generalization capability was validated on the public GC10-DET dataset. The proposed model achieved a 2.8% increase in mean average precision compared with the baseline detector, demonstrating that the proposed feature enhancement and fusion strategies are transferable to different industrial surface defect detection scenarios. These results indicate that the proposed framework effectively integrates accuracy improvement with computational reduction, which is essential for real-time industrial inspection systems.

A context-aware multi-scale fusion real-time detection model for forged-part surface defects is presented. Directional gradient fusion, adaptive contextual feature integration, and geometry-sensitive regression jointly enhance detection robustness under complex backgrounds and severe scale variation. Experimental results confirm that the proposed approach improves detection accuracy while reducing computational complexity and maintaining high inference speed. The model shows strong generalization capability and practical applicability, providing an effective solution for automated forged-part surface defect inspection in industrial environments.

Weak signal detection under complex backgrounds is a key technology in fields such as medical healthcare, aerospace engineering, and other related areas. The phase-locked amplifier (LIA) can effectively extract the sinusoidal signal of a specified frequency from noise, featuring advantages of a narrow passband and high signal-to-noise ratio (SNR). However, traditional lock-in amplifiers face multiple issues including insufficient hardware scalability and limited software flexibility. Especially in precision measurement tasks requiring real-time adjustment of modulation parameters, the limitations of their fixed hardware architecture and single control logic become increasingly prominent. To address these problems, this paper proposes a programmable lock-in amplifier system based on the LabVIEW FPGA architecture, which realizes the collaborative work of hardware and software.

In terms of hardware, the system takes the XILINX XC7A200T FPGA chip as the hardware core, and constructs a highly reliable hardware platform with the ALINX AX7202 backplane and AC7200 core board. Meanwhile, the ADC module AN9238 (12-bit resolution, 65 MSps sampling rate) and DAC module AD9767 (14-bit resolution, 125 MSps sampling rate) are selected to ensure high fidelity of signal acquisition and output.

In terms of software, with the LabVIEW graphical programming platform as the core, the system can flexibly configure the entire signal processing workflow. For the modulation and demodulation function: first, high-throughput function operations are innovatively adopted to realize waveform generation. This scheme is based on a 120 MHz clock-driven timing cycle structure, implements the accumulator function via a shift register, and dynamically generates waveform data using high-throughput multiplication, division, and sine function operations. This not only reduces FPGA resource consumption but also achieves flexible phase configuration and minimized data delay, laying the foundation for the dynamic optimization of subsequent modulation parameters. Second, in the phase-sensitive detection (PSD) module, weak signal demodulation and extraction are realized based on the mixing principle: the signal under test collected by the ADC is multiplied with the reference signal generated internally in the FPGA, and useful signals are initially separated from noise by leveraging the characteristics of coherent same-frequency signals and non-coherent noise. According to phase-locked amplification theory, when the frequency of the signal under test matches that of the reference signal with a fixed phase difference, multiplication generates a DC component and a frequency-doubled component. The DC component is proportional to the amplitude of the signal under test, providing a theoretical basis for signal extraction. Finally, in the filtering module, a cascaded design of a mean filter and Butterworth low-pass filter is adopted: the mean filter is configured with a 60-sample window, performing signal smoothing within a single-cycle timing loop to initially filter out high-frequency random noise; the subsequent 4th-order Butterworth filter adopts a cutoff frequency configuration strategy with a bandwidth margin of two to three times, further suppressing residual high-frequency components without signal distortion, thus achieving an optimal balance between SNR and response speed.

In the information transmission part, the UDP communication IP core of the host computer is invoked, which is interconnected with the physical pins and FIFO buffer in the FPGA through a four-wire handshaking mechanism. The host computer software is developed based on LabVIEW, featuring a comprehensive human-machine interface (HMI) that supports real-time configuration of core parameters (modulation frequency: 5 Hz–16 kHz adjustable; signal amplitude; phase: 0°–360° continuously adjustable; filter order) and provides a real-time waveform display function, significantly improving the real-time performance and stability of signal processing.

To verify the feasibility and practical value of the system, this paper deeply integrates the programmable lock-in amplifier with TDLAS technology to construct a high-precision trace oxygen detection system. A laser with a central wavelength of 760.59 nm and a Herriott-type multi-pass gas absorption cell are employed. First, a modulation amplitude optimization experiment is conducted, constructing a 0.3–1.8 V modulation coefficient gradient test under the conditions of 5 Hz scanning frequency, 16 kHz modulation frequency, 210° relative phase, and 200 Hz filter bandwidth. The results verify that the relationship between the second harmonic amplitude and modulation amplitude meets expectations, i.e., an optimal modulation amplitude exists. Second, concentration calibration and linearity verification experiments are performed: with a 5 Hz scanning frequency and 16 kHz laser modulation frequency, standard oxygen samples with concentrations ranging from 0.02% to 15% are prepared, and real-time detection and calibration of the second harmonic signals are conducted. The fitting correlation coefficient (R²) between oxygen concentration and the second harmonic signal is approximately 0.999, verifying the system’s feasibility and practicality. Finally, detection limit analysis is carried out, and calculations show that the system’s oxygen detection limit is 2.8×10⁻⁵.

This lock-in amplifier has the advantages of high integration, fast detection speed, and high SNR, making it suitable for on-site detection of trace oxygen. This design can exhibit significant advantages in spectral analysis scenarios requiring multi-parameter continuous adjustability and high real-time performance, promote the application and development of weak signal detection technology, enrich the implementation pathways of digital lock-in amplification technology, and provide a new design idea for the development of high-flexibility weak signal detection systems.

In the production process of printed circuit boards (PCBs), common defects such as mouse_bite, open_circuit, short, spur and spurious_copper are frequently encountered. These defects are characterized by their small size, color similarity to the background, and similar features across different defect types, which make detection challenging and prone to false positives and missed detections. Traditional inspection methods, including manual visual checks and conventional automated optical inspection (AOI), suffer from high error rates, limited adaptability, and dependency on manually selected features. Although deep learning approaches, particularly YOLO-series models, have improved detection robustness, they still face three core limitations in PCB defect inspection: insufficient feature extraction for tiny objects, poor balance between model lightweight and accuracy, and inadequate loss function design for small targets. To address the issues of limited samples, low efficiency, large parameter size, and low accuracy in detecting tiny-defects on PCBs, this paper proposes a lightweight detection model named PCB-YOLO based on the YOLOv8n.

First, the original high-resolution PCB images with diverse background colors are converted to grayscale to minimize color interference. To avoid losing tiny defects during resizing, a sliding-window slicing strategy is applied, generating 640 pixel×480 pixel sub-images with 100 pixels overlaps. Only slices containing defects are retained, effectively increasing the usable dataset while preserving small-object details. Second, considering the significant proportion of small target defects in PCBs, a Conv_SWS (SimAM with slicing) module is designed by integrating the SimAM attention mechanism and merging slicing operations. The input feature map is divided into local patches, each enhanced by SimAM to emphasize small-defect regions, before being recombined. This module replaces standard convolution in downsampling stages, enhancing multi-scale feature extraction without adding parameters. Third, the original C2f module in YOLOv8 is modified by introducing partial convolution (PConv) to comprehensively compute key feature channels, resulting in the design of the FBC2f module. This enhancement improves feature extraction capability while reducing the number of model parameters. Finally, the standard IoU loss is sensitive to minor localization errors and fails when boxes do not overlap. Thus, a combined loss using normalized Wasserstein distance (NWD) and IoU is proposed. NWD models bounding boxes as 2D Gaussian distributions and measures similarity via Wasserstein distance, providing stable gradients even for non-overlapping tiny objects. The final loss is a weighted sum of NWD and IoU, optimized with a balancing factor α=0.5. The model is trained and evaluated on an augmented PCB defect dataset containing 5172 images with 6282 defect instances. Performance is measured using mAP50, mAP50-95, parameters, FLOPs, and frame per second (FPS). Comparative experiments include ablation studies, attention mechanism comparisons, loss function analysis, and benchmarking against state-of-the-art detectors.

Experimental results demonstrate the effectiveness of PCB-YOLO: it achieves 95.6% mAP50 on the PCB defect dataset, outperforming YOLOv8n by 2.9%. It also reduces parameters by 23.6% (from 3.006 M to 2.302 M), FLOPs by 22.2% (from 8.1 G to 6.3 G), and model weight by 1.3 M. Moreover, ablation studies are finished to evaluate each proposed component contributes positively: the Conv_SWS module boosts recall by 5% and mAP50 by 2.4%; the FBC2f module reduces parameters by 23.5% with a moderate mAP50 trade-off. the NWD-IoU loss increases mAP50 by 2.3% and mAP50-95 by 2.8%. The full integration of all three modules achieves the best balance of accuracy and efficiency. Comparative evaluations demonstrate that PCB-YOLO outperforms Faster R-CNN, YOLOv5n, YOLOv10n, YOLOv11n, YOLOv12n, and the recently introduced lightweight PIC2f-YOLO in terms of both accuracy and model size. Additionally, it achieves a real-time inference speed of 160 f/s on GPU and 50 f/s on CPU, meeting the throughput demands of industrial applications. Moreover, PCB-YOLO also demonstrates strong generalization on external datasets, attaining mAP50 scores of 98.1% on the industrial aluminum surface defects dataset and 93.6% on the PKU-Market-PCB dataset, thereby verifying its robustness and transferability. Visual comparisons reveal that PCB-YOLO substantially decreases the rates of missed detections and false alarms relative to YOLOv8n, especially in the case of minor defects like shorts and mouse bites. In summary, the results validate that the Conv_SWS module effectively preserves small-object features during downsampling, the FBC2f module maintains representational capacity while reducing computation, and the NWD-IoU loss stabilizes training and improves localization for miniature targets.

This paper proposes PCB-YOLO, a lightweight and accurate detection model designed for tiny PCB defects. The approach integrates grayscale transformation and adaptive slicing during preprocessing to enhance input data quality. It introduces the Conv_SWS module to improve small-object feature retention and designs the parameter-efficient FBC2f module. Additionally, the NWD-IoU loss is adopted to refine the regression of minor defects. Together, these contributions address key challenges in PCB inspection, including limited sample availability, background variation, small defect size, and the accuracy–complexity trade-off. Experimental results show that PCB-YOLO achieves higher detection accuracy with fewer parameters and faster inference than current leading models. It also demonstrates strong generalization across diverse industrial defect datasets. The framework offers a practical, deployable solution for real-time, high-precision PCB quality inspection in manufacturing. Future work may explore further architectural optimizations and extend the method to other micro-defect detection domains.

In the context of the information age, massive data growth has posed severe challenges to traditional storage technologies such as magnetic storage and hard disks, which are limited by low storage density, high energy consumption, and short service life, failing to meet the long-term and low-cost storage demands of large-scale datasets. Holographic data storage, as a promising optical storage technology with advantages of high density, long lifespan, and high speed, has become a key solution to address this bottleneck. Deep learning can be used to combine the holographic storage system to demodulate the complex amplitude information. This paper focuses on the correlation mechanism between the diffraction distance, diffraction intensity image, and deep learning decoding in the holographic storage system, and explores the differences in the representation of complex amplitude information by intensity images under different diffraction distances and the influence patterns on the deep learning model.

To achieve the research objectives, firstly, a theoretical analysis was conducted on the relationship between the oversampling rate and the diffraction distance. The non-interference complex amplitude decoding holographic storage system was emphasized, and the optimal diffraction distance of 20 for the oversampling rate was explored under this system. A decoding framework based on deep learning was constructed, leveraging the powerful image feature extraction and mapping capabilities of the U-Net neural network. Intensity diffraction images corresponding to different diffraction distances were collected, which contained the implicit phase and amplitude information of the complex amplitude data pages. These intensity diffraction images were used to train the U-Net model to explore the intrinsic correlation between the diffraction distance and the feature description of the complex amplitude. Additionally, by adjusting the oversampling rate parameters, the relationship between the oversampling rate and the diffraction distance was analyzed, and the symbol error rate (SER) decoding performance indicators were statistically compared under different experimental conditions to verify the relationship between the oversampling rate and the diffraction distance.

Firstly, it clarifies the correlation rule between oversampling rate and diffraction distance: the higher the oversampling rate, the longer the diffraction distance required to obtain an intensity diffraction pattern with a specific contrast. This is because a higher oversampling rate leads to a more compact distribution of spectral information, thus requiring a longer diffraction distance to separate the complex amplitude features hidden in the intensity image. Secondly, under specific conditions (oversampling rate of 20 and diffraction distance of 1.4 millimeters), the complex amplitude features carried by the intensity diffraction pattern match highly with the feature extraction mechanism of the U-Net neural network. This matching effect improves the efficiency of feature mapping of phase and amplitude information during model training, and makes the SER approach 0 in the early stages of iteration, which is superior to the decoding performance under other diffraction distance conditions. Finally, the oversampling rates of 10 and 30 correspond to the optimal diffraction distances of 0.4 mm and 3.1 mm respectively, further verifying the correlation rule between oversampling rate and diffraction distance elucidated in this paper.

This paper focuses on the core correlations between the diffraction distance, intensity diffraction pattern, and the specific Unet deep neural network decoding in the holographic storage system. It particularly explores the intrinsic relationship between the oversampling rate and the diffraction distance, and conducts quantitative analysis on the optimal diffraction distance under specific parameter combinations. The correlation rule between the oversampling rate and the diffraction distance is clarified: the higher the oversampling rate, the greater the diffraction distance required to obtain the specific contrast intensity diffraction pattern. This rule and the quantitative results of the optimal diffraction distance provide specific references for the selection of diffraction distance and parameter optimization in similar holographic storage systems, and lay the foundation for subsequent research on different encoding schemes and storage media.