The unstructured road environment in open-pit mining areas is characterized by complex topography, variable lighting conditions, and irregular boundaries. In autonomous driving scenarios for mining trucks, particularly under medium-to-long-distance viewing perspectives, critical obstacles such as ruts and puddles often appear as low-resolution, small-scale targets. Conventional convolutional neural networks typically employ strided convolutions or pooling operations for downsampling, which inevitably leads to the loss of fine-grained texture and boundary information. Consequently, existing segmentation methods frequently suffer from missed detections of small targets and blurred segmentation edges, posing severe safety risks to the path planning of unmanned vehicles. To address these challenges, this study aimed to develop a high-precision, real-time instance segmentation model, named SCG-YOLO, specifically optimized for the drivable areas of open-pit mine roads. The primary goal was to enhance the perception of low-resolution details and irregular boundaries while strictly adhering to the computational constraints of onboard embedded devices.

The proposed SCG-YOLO model was constructed based on the YOLOv8-seg architecture, integrating three targeted improvement modules to handle the unique constraints of mining environments.

First, a space-to-depth and deformable convolution (SDC) module was designed and embedded into the feature extraction network. To mitigate the information loss associated with traditional downsampling, the Space-to-Depth layer decoupled and reorganized the spatial dimensions of input feature maps into depth dimensions. This operation effectively preserved fine-grained information, such as the shallow textures of ruts and the irregular edges of puddles, which are typically discarded by standard pooling layers. Furthermore, Deformable Convolution layers were incorporated within the SDC module. Unlike standard convolutions with fixed geometric structures, deformable convolutions learned offset fields to dynamically adjust the sampling grid. This capability allowed the network to adaptively model the non-rigid deformations of targets on unstructured road surfaces, ensuring accurate feature alignment for distorted shapes.

Second, a coordinate attention (CA) mechanism was introduced into the network neck. Distinct from global average pooling which compresses spatial information, the CA mechanism performed feature encoding along two spatial directions (horizontal and vertical) respectively. This process captured long-range dependencies while preserving precise positional information. By re-weighting the feature maps, the mechanism enhanced the network's focus on key drivable regions and suppressed interference from complex background factors common in mining areas, such as ore piles, dust, and equipment shadows.

Third, to address the increased computational overhead introduced by the SDC and CA modules, a Ghost module was adopted for the lightweight reconstruction of the neck layer. Standard Convolution and C2f modules were replaced with GhostConv and C3Ghost modules. The Ghost module generated a portion of the feature maps using standard convolutions and then produced redundant features via computationally inexpensive linear operations. This strategy significantly reduced the number of parameters and floating point operations (FLOPs) without compromising the representational power of the model, thereby serving as an engineering constraint to ensure real-time performance.

Experimental validation was conducted using a self-collected dataset of open-pit mine roads, comprising 565 images covering various seasons and lighting conditions. The dataset included three annotated categories: drivable areas, ruts, and puddles. The proposed SCG-YOLO model demonstrated superior performance across multiple evaluation metrics.

Quantitative analysis revealed that the model achieved a mean average precision (mAP50) of 82.6%, representing a substantial increase of 11.7% compared to the baseline YOLOv8-seg model. In terms of inference speed, the model maintained a frame rate of 425.31 f/s, satisfying the stringent real-time requirements of autonomous mining operations.

Ablation studies provided deep insights into the contribution of each component. The integration of the SDC module alone yielded a 9.4% improvement in mAP50, validating its efficacy in recovering fine-grained details lost during downsampling. The addition of the CA mechanism contributed a further 2.2% increase, confirming its ability to filter background noise and sharpen focus on relevant features. The implementation of the Ghost module successfully reduced the parameter count to 11.2 M, offsetting the computational cost of the accuracy-enhancing modules.

Comparative experiments conducted with various instance segmentation algorithms, including Mask R-CNN, YOLACT, FastInst, Mask2Former, and various YOLO iterations (v5, v6, v11), highlight the advantages of SCG-YOLO. While heavy models like Mask2Former offered competitive accuracy, their excessive computational demands (286.5 G FLOPs) rendered them unsuitable for edge deployment. Conversely, lighter models like YOLACT failed to achieve sufficient segmentation precision for small targets. SCG-YOLO achieved the optimal balance between accuracy and efficiency.

Visual qualitative analysis showed that the model produced significantly sharper segmentation masks. In scenarios with dusty backgrounds or water-filled potholes, SCG-YOLO accurately delineated boundaries that appeared blurred in baseline predictions. Generalization tests performed on the public RUGD dataset further demonstrated the model's robustness, where it outperformed the baseline by 5.2% in mAP50, proving its adaptability to diverse unstructured terrains beyond the specific training site.

The study presents a robust solution for environmental perception in open-pit mining based on spatial depth conversion. The investigation confirms that the proposed SCG-YOLO model effectively resolves the core issues of fine-grained feature loss and edge blurring associated with low-resolution imaging of small targets in unstructured road environments. The synergistic combination of the SDC module and CA mechanism significantly enhances the model's ability to perceive rut textures and puddle boundaries under complex conditions. Simultaneously, the lightweight design utilizing Ghost modules ensures that the high computational efficiency required for autonomous driving systems is maintained. The model achieves a high segmentation precision of 82.6% mAP50 while running at speeds exceeding 400 f/s, providing a viable and safe technical foundation for the deployment of unmanned mining trucks. Future research directions include expanding the dataset to encompass extreme weather conditions and multi-vehicle interaction scenarios to further improve the model's generalization capabilities in dynamic mining environments.

Semantic segmentation of urban street scenes plays a central role in perception systems for autonomous driving, where reliable pixel-level understanding of roads, vehicles, infrastructure, and vulnerable road users is essential for safe planning and control. Existing real-time networks still struggle to delineate object boundaries under occlusions, scale variation, cluttered backgrounds, and complex lighting, which leads to fragmented objects, blurred contours, and missed small instances such as poles, traffic signs, riders, and pedestrians. Many architectures emphasize global context or throughput, yet fail to model fine edges explicitly and therefore sacrifice boundary quality. This study aimed to enhance boundary precision and small-object segmentation while preserving high inference speed, through the design of a dual-branch network that exploits explicit boundary refinement, adaptive receptive fields, and cross-branch information propagation for street-scene understanding.

A dual-branch boundary refinement network (BRNet) was developed for street-scene semantic segmentation. The architecture contained a high-resolution branch with three adaptive sampling convolution (ASC) blocks that preserved fine spatial structure and adjusted receptive fields dynamically to complex object geometry. A low-resolution branch integrated a Boundary Refinement Module (BRM) that expanded the effective kernel to a 5×5 region with double convolutions and employed lightweight channel and spatial attention to strengthen edge features and suppress background noise near object contours. A feature propagation module (FPM) linked the two branches through progressive cross-scale fusion, compressing channels and performing downsampling or upsampling so that enhanced semantic and boundary cues from the low-resolution path flowed back to the high-resolution path and complemented local details. This design formed a collaborative boundary optimization mechanism, where both branches contributed to a consistent refinement of contours at different resolutions. Supervision signals were attached to multiple stages, and a hybrid loss combined cross-entropy with Lovasz-Softmax loss to optimize intersection-over-union directly and alleviate the class-imbalance characteristic of street-scene datasets, in which background and road pixels dominate rare categories. The network was trained and evaluated on the Cityscapes benchmark with fine annotations. The training pipeline adopted stochastic gradient descent with a poly learning-rate schedule, and used common data augmentation operations such as random scaling, cropping, color jittering, and horizontal flipping to increase robustness to viewpoint and appearance variations. Implementation relied on a standard deep learning framework with synchronized batch normalization and mixed-precision computation to stabilize optimization and accelerate training. Performance was assessed by mean intersection over union (mIoU), mean pixel accuracy (mPA), and frames per second (FPS) to capture both segmentation quality and computational efficiency. Additional experiments on the WildDash2 and BDD100K datasets examined robustness under distribution shifts that involved different cities, illumination conditions, camera setups, and adverse weather.

BRNet achieved an mIoU of 81.43% on the Cityscapes test set and delivered a frame rate of 67 f/s on a single GPU, which demonstrated a favorable balance between accuracy and real-time performance. The method outperformed the bilateral segmentation network BiSeNetV2 by 5.69 percentage points in mIoU and exceeded other multi-branch or lightweight baselines such as DDRNet, LMIINet, and SegNeXt on most evaluation metrics, while remaining close to transformer-based EfficientViT in accuracy with noticeably lower computational cost. Quantitative analysis of small-object categories showed consistent gains over BiSeNetV2, DDRNet, and PIDNet for poles, traffic lights, traffic signs, pedestrians, motorcycles, and bicycles, confirming that explicit boundary modeling and adaptive sampling improved recognition of thin structures and distant targets. A separate comparison with models that integrated Mamba or transformer blocks indicated that BRNet obtained comparable segmentation quality with fewer parameters and reduced latency, because boundary attention was introduced through compact convolutional modules instead of heavy global self-attention. Cross-dataset tests further validated the generalization ability of BRNet. On the BDD100K dataset the network reached an mIoU of 43.18% and maintained a smaller performance drop than competing models when transferred from Cityscapes to unfamiliar urban environments that contained night scenes, rain, and complex traffic patterns. On the WildDash2 benchmark BRNet preserved coherent road layouts and object masks in images with motion blur, low visibility, and camera distortions. Qualitative visualizations on WildDash2 and BDD100K revealed clearer object contours, better separation of overlapping vehicles and pedestrians, and more complete road and sidewalk regions, especially in regions with shadows, glare, or motion blur. Class activation maps indicated that the boundary refinement mechanism guided attention to compact regions along object outlines, which helped preserve structural integrity, reduced boundary ambiguity, and mitigated confusion between foreground and background.

The dual-branch BRNet with collaborative boundary refinement improved semantic segmentation of urban street scenes, especially in boundary precision and small-object delineation, while keeping real-time inference capability suitable for on-board deployment. The high-resolution branch with adaptive sampling captured detailed spatial layouts and local textures, and the low-resolution branch with the boundary refinement module enhanced semantic discrimination along contours; the feature propagation module linked these components into a continuous optimization loop for boundary representation across scales. The hybrid loss function reduced the impact of class imbalance and supported stable optimization toward high IoU scores without sacrificing pixel-wise accuracy. Experiments on Cityscapes, WildDash2, and BDD100K confirmed strong generalization across datasets and robustness to challenging imaging conditions such as adverse weather and poor illumination. These characteristics make BRNet a competitive candidate for perception stacks of autonomous driving and advanced driver-assistance systems that require accurate, efficient, and reliable understanding of complex urban environments. Future work will refine the architecture and compression strategy to further reduce latency and memory footprint, and will explore integration with panoptic segmentation, instance-level boundary detection, temporal consistency modeling, and multi-task learning frameworks to obtain richer and more coherent scene understanding from a single unified model.

Image classification, a foundational core task in computer vision, supports diverse critical applications including clinical medical diagnosis, intelligent security surveillance, remote sensing image interpretation, and autonomous driving scene recognition. Deep learning models such as ResNet, DenseNet, and ViT have notably advanced classification performance in recent years, but traditional networks still bear inherent flaws that restrict their practical applicability in complex real-world scenarios. ResNet, a widely adopted architecture for image classification tasks due to its gradient preservation advantage, lacks sufficient channel feature interaction, which directly impedes the effective selection of key discriminative features. Its initial 7×7 convolution and max-pooling layers trigger excessive downsampling at the early feature extraction stage, resulting in the loss of fine-grained spatial details that are vital for accurate classification of complex images with overlapping objects or subtle differences. Fixed channel grouping in traditional convolutions further isolates cross-channel feature information, limiting the model’s ability to integrate multi-dimensional feature cues and capture inter-channel correlations. Existing attention mechanisms like SE and CBAM only focus on single dimensions—either channel or spatial—failing to achieve dynamic and comprehensive feature calibration varying image contents. This study addresses these critical limitations, enhances cross-channel information interaction, strengthens feature representation and key feature extraction capabilities, and ultimately improves the overall classification performance and generalization ability of the model.

To achieve this objective, a pooling residual image classification network with dynamic channel feature calibration (DCPRNet) was proposed, which is optimized and improved based on the ResNet-34 architecture to address its inherent deficiencies. Three core targeted improvements were integrated into the network to overcome the drawbacks of traditional models while maintaining computational feasibility. First, the initial layer was optimized: the 7×7 convolution was replaced with a 3×3 convolution to reduce parameter complexity by nearly half while retaining robust feature extraction capability, and the max-pooling layer was removed to fully preserve fine-grained spatial information for subsequent feature processing, fusion, and calibration. Second, a dynamic channel feature calibration (DCFC) module was designed, which combines a random subgroup shuffle module and chunked weighted attention (CWA) mechanism. The random subgroup shuffle module breaks the isolation between channel groups via dual-layer grouping and random shuffling, effectively promoting cross-group information interaction and dynamic channel communication without introducing excessive computation. The CWA mechanism calibrates the weights of feature chunks through feature chunking, local importance score generation, mean pooling-based global information aggregation, and sigmoid normalization, thereby suppressing redundant background information and enhancing the contribution of critical feature regions related to target classification. Third, an average pooling-based residual module was added to the DCFC branch, using 1×1 convolution and 2×2 average pooling instead of traditional stride convolution to realize low-cost smooth downsampling, reduce feature distortion, and better preservation of effective feature information.

Comprehensive experiments were conducted on five representative datasets (CIFAR-10, CIFAR-100, SVHN, Imagenette, Imagewoof) that cover simple object, complex category, low-resolution, and real-world natural image scenarios, ensuring the model’s performance is validated under diverse conditions. DCPRNet achieved classification accuracies of 96.41%, 80.36%, 96.97%, 91.59%, and 80.61% on these datasets respectively, showing stable and reliable performance across different data types, complexity levels, and feature distributions. Ablation studies further confirmed that input layer optimization, the DCFC module, and the average pooling residual module each boost model performance significantly, and the synergistic effects among the three components yield the optimal overall result by complementing each other’s advantages. Comparative experiments against 12 mainstream models, including ResNet-34, DenseNet-121, ViT-B/16, and MobileNet-V2, demonstrated DCPRNet’s consistent superiority: it achieved a 2.15% accuracy gain over the baseline ResNet-34 on the CIFAR-100 dataset and a 4.39% improvement on the ImageNet dataset, while maintaining comparable computational cost. Heatmap-based visual analysis further verified that DCPRNet can accurately focus on critical fine-grained features of target objects, effectively avoiding interference from irrelevant background regions and exhibiting excellent feature calibration capability. These results confirm that the proposed improvements effectively enhance cross-channel information interaction, resolve the inherent limitations of traditional networks, and strengthen the model’s ability to extract discriminative features.

DCPRNet was successfully developed by optimizing the ResNet-34 architecture and introducing the DCFC module and average pooling residual module. The model effectively enhances cross-channel information interaction, significantly strengthens feature representation ability and key feature extraction efficiency, and achieves a balanced trade-off between classification accuracy, computational cost, and model robustness. Experimental results and visual analysis fully validate its superiority over mainstream models on multiple datasets, confirming its good robustness and adaptability to different image scenarios and data distributions. By addressing the core flaws of traditional image classification networks—such as insufficient channel interaction, excessive feature loss during downsampling, and limited feature calibration capability—DCPRNet provides a reliable technical solution for image classification tasks in multiple fields including medical imaging, intelligent surveillance, and remote sensing detection. It also offers valuable theoretical and practical reference for the further optimization of deep learning-based image classification models, laying a solid foundation for the development of more efficient, accurate, and lightweight computer vision algorithms suitable for edge computing devices.

Traditional diffraction lenses can manipulate wavefronts by adjusting the refractive index of materials and surface profiles of devices, which are widely applied in optical imaging. However, such imaging capabilities are usually constrained by limitations of low focusing efficiency, a single operating wavelength, and chromatic aberration. To suppress chromatic aberration and improve image quality, it is commonly used by combining positive and negative refractive index lenses. Nevertheless, this scheme hinders the miniaturization and integration of optical systems. Although monolithic refractive-diffractive hybrid lenses provide an alternative method to enhance system integration for achromatic performance, complex structural design restricts their further applications. Recently, metasurfaces offer another methodology for realization of broadband achromatic, but the high processing precision in micro-nanofabrication is always required. Consequently, there is still a great challenge for designing diffractive lenses with simple structure, low processing difficulty, and achromatism.

For conventional direct binary search (DBS) algorithm to optimize diffraction lenses, the random initial structures are commonly used, whereas it commonly leads to slow convergence, sensitivity to the starting state. Here, a modified DBS algorithm is proposed for designing three-wavelength multi-order achromatic diffraction microlenses by directly tuning the discrete surface-relief heights. Utilizing the radially symmetric initial height distribution, such approach can reduce the randomness and improve computational stability and corresponding main procedures are as follows: First, the ideal continuous phase distribution of achromatic focusing lens with a parabolic phase profile can be generated; Second, such continuous phase is quantized into discrete phase corresponding to different step-heights for optimizing multilevel diffractive lens by iterative perturbations of local phase; Third, a radially symmetric initial structure determined by modulo function is introduced to replace the conventional random initialization, and then the step heights corresponding to different wavelengths are alternately distributed at distinct positions. Finally, a figure-of-merit (FOM) is adopted to evaluate the performance of the designed lens and guides the iterative updates of the step-height distributions. As a proof of concept, the 8-step and 16-step diffraction lenses (with a diameter, focal length, and numerical aperture of 560 μm, 10 mm, and 0.028, respectively) are designed and analyzed both theoretically and through simulation.

For pre-optimization, the average focusing efficiencies of the diffractive lenses with 8 phase levels (L=8) and 16 phase levels (L=16) are 77.04% and 78.82% at the wavelengths of 1.064 μm, 1.300 μm, and 1.550 μm, respectively. Therefore, the average focusing efficiencies of both lenses are less than 79% affected by chromatic aberration in theory. To improve the focusing efficiency and suppress chromatic aberration, a modified DBS algorithm is employed to optimize the step-height distribution. For post-optimization, the focusing efficiencies of diffraction lens (L=8) can reach 87.92%, 94.32%, and 90.73% (the efficiency is 90.99%) at 1.064 μm, 1.30 μm, and 1.55 μm after 4000 iterations, respectively, which is improved 13.95% compared to pre-optimization in theory. When L is changed to 16, the efficiencies exceed 90.01%, 96.63%, and 90.73% (at 1.064 μm, 1.300 μm, and 1.550 μm) respectively, and corresponding average value surpasses 92.46% (the improvement is 13.64% compared to pre-optimization) under the same conditions. These results indicate that the proposed achromatic diffraction lens maintains higher performance with the proposed modified DBS algorithm in theory. To further validate the achromatic performance, numerical simulations are conducted using CST microwave studio. For the diffraction lens with L=8 and 16, the focusing efficiencies at 1.064 μm are 87.17% and 89.53% in simulation, respectively. Moreover, it can be increased to 89.10% and 96.24% at 1.30 μm, while the efficiencies are 89.10% and 89.08% at 1.55 μm, respectively. Therefore, corresponding average efficiencies can reach 90.09% (L=8) and 91.62% (L=16), respectively. Note that the slight difference between theoretical prediction and numerical simulation may be attributed to the ideal assumptions (neglected the interactions between structures) in the theory. Furthermore, the point spread functions (PSFs) at the focal plane for different wavelengths are employed to characterize the achromatic performance of the designed diffraction lens. The simulated full width at half maximums (FWHM) are 25.8 μm, 26.9 μm, and 33.4 μm (L=8) and 26.9 μm, 26.9 μm, and 33.4 μm (L=16) at 1.064 μm, 1.300 μm, and 1.550 μm, respectively, which are agree well with the theoretical results of 21.8 μm, 26.6 μm, and 32.2 μm and close to their diffraction limits of 19.00 μm, 23.22 μm, and 27.69 μm at different wavelengths. Totally, both the 8-step and 16-step diffraction lenses possess an average focusing efficiency of ~79% (at 1.064 μm, 1.30 μm, and 1.55 μm) before optimization. However, the average efficiencies can exceed 90.99% and 92.46%, (improvements of 13.95% and 13.64%), respectively, based on the improved DBS algorithm (after 4000 iterations) in theory. Moreover, it has also demonstrated that the FWHM of the 8-step and 16-step diffraction lenses can reach the diffraction limit at three wavelengths, exhibiting high imaging performance.

Here, a new methodology is proposed for achieving three-wavelength achromatism, and may offer advantages in reduction of device complexity, requirements of machining accuracy, and the manufacturing costs. These results may be further extended to diverse applications, such as three-wavelength imaging systems, optical computing, lightweight collimators, optical communications, biomedical imaging, and laser processing.

A sparse sampling mode set including key views and non-key views is constructed via the core strategy of sparse coding and reconstruction, providing fundamental support for the coding of wide-FOV light field images with variable bitrates. The specific implementation scheme is as follows: Firstly, the multi-dimensional complexity features of light field images are extracted, and an adaptive selection scheme for sparse sampling modes is designed combined with target bitrate constraints. Leveraging the K-nearest neighbor (KNN) algorithm, rapid matching of the optimal mode is enabled, allowing identification of the best sparse sampling strategy without traversing the entire mode set. On this basis, predictive relationships among key views are established, providing a more robust basis for subsequent coding and compression processes. Secondly, to address uneven view quality caused by traditional bitrate allocation methods, a frame-level bitrate pre-allocation strategy is designed according to the inter-frame spatial distance between the current encoded frame and its reference frames within the key view set. Dynamic adjustment of bitrate allocation weights effectively improves the angular quality consistency of the encoded views. Then, based on the I-frames in the key view set, the saliency detection and depth estimation techniques are integrated to generate largest coding unit (LCU)-level saliency weights. Utilizing the angular correlation of light field images, these weights are propagated to all key views along the angular dimension. Furthermore, saliency weights are combined with image regional complexity to develop a fine-grained LCU-level content-aware bitrate allocation strategy, achieving more accurate bitrate allocation. Finally, at the client side, key views are decoded according to the transmitted optimal sparse sampling mode, and non-key views are reconstructed using the decoded key views with a deep learning-based view synthesis network, ultimately completing the reconstruction of the wide-FOV light field image.

In this paper, the core strategy of sparse coding and reconstruction is adopted to construct a sparse sampling mode set that includes key views and non-key views, thereby providing fundamental support for the coding of wide-FOV light field images with variable bitrates. The specific implementation scheme is as follows: Firstly, the multi-dimensional complexity features of light field images are extracted and an adaptive selection scheme is designed for sparse sampling modes in combination with target bitrate constraints. Leveraging the K-nearest neighbor (KNN) algorithm, this scheme enables the rapid matching of the optimal mode, which allows the identification of the best sparse sampling strategy without traversing the entire mode set. On this basis, the predictive relationships among key views are established, providing a more robust basis for subsequent coding and compression processes. Secondly, aiming to address the issue of uneven view quality caused by traditional bitrate allocation methods, a frame-level bitrate pre-allocation strategy is designed according to the inter-frame spatial distance between the current encoded frame and its reference frames within the key view set. By dynamically adjusting the bitrate allocation weights, this strategy effectively improves the angular quality consistency of the encoded views. Then, based on the I-frames in the key view set, the saliency detection and depth estimation techniques are integrated to generate LCU (largest coding unit)-level saliency weights. Utilizing the angular correlation of light field images, these weights are propagated to all key views along the angular dimension. Furthermore, the saliency weights are combined with the regional complexity of images to develop a fine-grained LCU-level content-aware bitrate allocation strategy, achieving more accurate bitrate allocation. Finally, at the client side, the key views are decoded according to the transmitted optimal sparse sampling mode, and the non-key views are reconstructed using the decoded key views with a deep learning-based view synthesis network, thus ultimately completing the reconstruction of the wide-FOV light field image.

The proposed adaptive sparse sampling mode selection scheme, leveraging multi-dimensional feature fusion and deep learning prediction mechanisms, significantly reduces the computational complexity of mode selection and avoids exorbitant computational costs incurred by the traditional brute force search method. It rapidly matches diverse target bitrates and image content features to select the optimal or sub-optimal sparse coding mode. The designed two-level bitrate allocation strategy, combining frame-level pre-allocation and LCU-level perceptual allocation, effectively mitigates uneven angular view quality while maintaining compatibility with existing video coding standards, further enhancing the angular quality consistency of wide-FOV light field images. Experimental results show that, compared with existing representative methods, the method achieves an average bitrate saving of 45.284% and a peak signal-to-noise ratio (PSNR) improvement of 1.641 dB, with the maximum random access penalty of light field images limited to only 20.29%. These results indicate that the method not only improves coding efficiency but also ensures favorable random access performance, satisfying the demand for rapid view retrieval in practical applications.

The proposed wide-FOV light field image coding method based on target bitrate and image content awareness effectively adapts to the coding requirements of variable bitrates and uncertain image content. It achieves satisfactory results in reducing sampling mode selection complexity and improving network transmission adaptability. Meanwhile, it fulfills multiple objectives including bitrate saving, image quality enhancement, angular quality consistency improvement, and random access capability optimization, demonstrating superior overall coding performance and favorable practical application value.

Unmanned aerial vehicle (UAV) infrared target detection faces significant challenges including limited computational resources, strong background interference, and multi-scale target variations. Infrared imaging, compared with visible light imaging, provides all-day and all-weather working capabilities with strong anti-interference performance. However, due to flight altitude and sensor limitations, UAV-acquired infrared images commonly suffer from low spatial resolution, poor signal-to-noise ratio, and large target scale spans. In long-distance shooting conditions, targets often exhibit multi-scale distributions, with small targets of several pixels and medium-to-large targets covering tens of pixels coexisting in the same scene. Traditional detection methods tend to produce high false positive rates and severe missed detection in complex backgrounds. To address these issues, this paper proposes a multi-scale infrared target detection algorithm based on YOLO, named YOLO-MITD. The algorithm aims to achieve high detection accuracy while maintaining low computational complexity, making it suitable for deployment on resource-constrained UAV platforms.

The proposed YOLO-MITD algorithm introduces four core modules to enhance detection performance. First, the ghost-enhanced cross-stage fusion module (GCFM) is designed by replacing standard convolutions in the original C2f module with GhostConv lightweight convolutions. GhostConv adopts a two-stage feature generation mechanism that uses standard convolution to extract essential features and then applies cheap transformations through depthwise separable convolution to generate phantom features, which significantly reduces model parameters while enhancing multi-scale feature extraction capabilities. The theoretical acceleration ratio approximately equals the redundancy factor s. Second, the efficient multi-scale interaction module (EMIM) incorporates the efficient multi-scale attention (EMA) mechanism into each bottleneck of the C2f structure. EMA adopts a "feature grouping plus cross-spatial learning" strategy to achieve one-time joint recalibration of channel and spatial information without dimensionality reduction in the channel dimension. The feature grouping number G is dynamically adjusted based on the channel count C, ensuring efficient feature representation while controlling computational complexity. Third, the adaptive spatial feature fusion head (ASFFH) is developed based on the adaptive spatial feature fusion (ASFF) mechanism, which consists of three steps: feature rescaling through interpolation or strided convolution to align spatial resolutions, adaptive fusion through softmax-normalized spatial weights learned by the network, and consistency guarantee to automatically suppress gradients from conflicting regions. This module dynamically allocates fusion weights for multi-scale features and suppresses feature conflicts, effectively improving scale consistency in target detection. Fourth, the adaptive threshold focal loss (ATFL) function is constructed with a piecewise weighting mechanism that dynamically adjusts the loss curvature. For high-confidence samples (p_t > 0.5), exponential decay mechanism compresses loss weights to low levels; for low-confidence samples (p_t $\leqslant $ 0.5), explosive growth mechanism amplifies loss weights to accelerate gradient backpropagation. The threshold parameter lambda is adaptively adjusted based on the exponential moving average of prediction confidence across training batches, thereby alleviating class imbalance issues during training.

Extensive experiments were conducted on the HIT-UAV dataset, which is the first publicly available UAV infrared thermal imaging target detection dataset for high-altitude scenes. The dataset contains 2898 UAV infrared aerial images covering multiple real-world scenarios including schools, parking lots, roads, and playgrounds, under both daytime and nighttime conditions. Notably, the target size distribution exhibits significant diversity, with small targets accounting for approximately 70%, medium targets about 28%, and large targets only 2% of the dataset. The proposed YOLO-MITD algorithm achieved an mAP50 of 86.7%, representing a 6.1% improvement over the baseline YOLOv8n. The precision and recall reached 89.7% and 83.3%, respectively, with increases of 5.6% and 8.3% compared to YOLOv8n. Despite these significant performance gains, the computational overhead increased by only 0.6 GFLOPs and the parameter count was controlled at 3.895 M, achieving an excellent balance between accuracy and complexity. Ablation experiments demonstrated the effectiveness of each module: GCFM reduced parameters by 0.485 M and FLOPs by 6.5 G while improving mAP50 by 1.8 percentage points; EMIM improved mAP50 by 4.2 percentage points with the most significant accuracy enhancement; ASFFH enhanced mAP50 by 2.6% but increased FLOPs by 2.2 G; ATFL improved mAP50 by 3.0% without increasing model complexity. The complete model combining all four modules achieved optimal performance, verifying the complementary and synergistic nature of the modules. Comparative experiments against mainstream models including RT-DETR-l, YOLO-SRMX, YOLOv8n, YOLOv8s, YOLOv10n, YOLOv10s, YOLOv11n, and YOLOv12n showed that YOLO-MITD achieves superior detection performance while maintaining lightweight characteristics. Specifically, compared with RT-DETR-l and YOLO-SRMX, mAP50 improved by 6.6% and 3.9%, respectively. Cross-scene generalization experiments on the CTIR and SIRST datasets further validated the algorithm's robustness, with mAP50 reaching 85.4% and 67.2%, respectively, significantly outperforming existing methods. Visualization results demonstrated that the proposed method reduces missed and false detections, particularly in complex backgrounds and for targets with incomplete contours.

This paper presents an improved YOLOv8n-based algorithm for multi-scale infrared target detection in UAV applications. The GCFM module achieves lightweight feature extraction through GhostConv by replacing standard convolutions with cheap operations, maintaining feature expressiveness while reducing computational burden. The EMIM module enhances feature representation via the EMA attention mechanism, which captures both channel and spatial dependencies through cross-dimensional interaction without dimensionality reduction. The ASFFH module improves multi-scale fusion through adaptive weight allocation, effectively resolving feature pyramid inconsistencies and suppressing gradient conflicts. The ATFL loss function optimizes the training process through dynamic threshold adjustment, applying different modulation strategies to samples based on confidence levels and alleviating class imbalance issues. The experimental results demonstrate that the proposed algorithm strikes an excellent balance between detection accuracy and model complexity, with strong cross-scene generalization capability. The method achieves state-of-the-art performance on multiple infrared datasets while maintaining low computational requirements, making it well-suited for real-time infrared target detection tasks on resource-constrained UAV platforms. Future work will focus on hardware deployment optimization, model quantization and pruning techniques, and further improvements to real-time performance and energy efficiency to adapt to more demanding engineering application requirements.

Narrow-lumen organs, including the urethra, bladder, esophagus, gastrointestinal tract, and blood vessels, are highly susceptible to pathological remodeling processes associated with inflammation, fibrosis, neovascularization, and microcirculatory dysfunction. Early-stage diseases occurring in these organs, such as urethral stricture, bladder cancer, esophageal lesions, and vascular abnormalities, are often accompanied by subtle structural alterations and local hemodynamic changes before obvious morphological narrowing becomes clinically detectable. Therefore, high-resolution visualization of tissue microstructure and microvascular perfusion is of great importance for early diagnosis, disease monitoring, and treatment evaluation. Conventional imaging modalities, including white-light endoscopy, ultrasound endoscopy, retrograde urethrography, and fluorescence imaging, still face limitations in simultaneously assessing tissue morphology and microcirculation. Endoscopic optical coherence tomography (OCT) provides micron-scale cross-sectional imaging of biological tissues, while optical coherence tomography angiography (OCTA) further enables label-free visualization of blood flow and microvascular networks. Compared with distal motor-driven probes, proximally controlled endoscopic OCT probes possess advantages in miniaturization, structural simplicity, and mechanical flexibility, making them more suitable for narrow and deformable luminal organs. However, proximally controlled endoscopic imaging systems are highly susceptible to non-uniform rotational distortion (NURD), rotational instability, and motion artifacts, which significantly degrade OCTA image quality and vascular visualization performance. In this study, a proximally controlled endoscopic OCT/OCTA system was developed and optimized for stable multi-organ in vivo imaging and pathological assessment in complex luminal environments.

A swept-source OCT/OCTA imaging system centered at 1310 nm was employed in this study. The system achieved an axial resolution of 7.46 μm and a lateral resolution of approximately 15 μm. A side-viewing proximally controlled endoscopic probe was designed using a single-mode optical fiber, a GRIN lens, and a micro-reflector to achieve circumferential luminal imaging. To improve OCTA imaging stability under rotational scanning conditions, a complementary NURD correction strategy combined with a helical B-scan acquisition scheme was implemented to suppress rotational artifacts and improve inter-frame registration accuracy. Optical simulations were further performed using ZEMAX software to optimize probe focusing performance, working distance, and assembly tolerance.

To validate imaging capability, multilayer adhesive tape samples were first imaged to evaluate structural resolution and imaging stability. Subsequently, in vivo OCT/OCTA imaging experiments were conducted across multiple organs and animal models, including rat rectum and esophagus imaging, mouse bladder imaging, and rabbit urethra and abdominal aorta imaging, covering digestive, urinary, and vascular systems. Furthermore, pathological imaging studies were performed using a rabbit urethral stricture model induced by electrocautery thermal injury and a C57 mouse bladder cancer model. Structural OCT images and vascular OCTA datasets were acquired simultaneously to evaluate disease-associated microstructural remodeling and microcirculatory alterations.

The optimized proximally controlled endoscopic OCT/OCTA system demonstrated stable imaging performance in multiple narrow-lumen organs with different anatomical structures and motion characteristics. Imaging results of multilayer adhesive tape samples clearly resolved layered microstructures, validating the optical focusing capability and structural imaging performance of the designed probe.

In vivo imaging experiments further demonstrated that the system could reliably visualize multilayer tissue architectures, including mucosal, submucosal, and muscular layers, in the rat rectum and esophagus. OCTA imaging successfully revealed superficial microvascular networks distributed within the mucosal and submucosal regions. In mouse bladder imaging, clear structural delineation of bladder wall layers and vascular distribution was achieved despite bladder deformation and motion. Imaging of rabbit abdominal aorta further demonstrated the capability of the system for vascular lumen imaging under pulsatile conditions.

For pathological imaging applications, rabbit urethral stricture tissues exhibited obvious structural remodeling characterized by urethral wall thickening, disordered layered morphology, and localized arc-shaped hyperreflective bands associated with fibrosis formation. Correspondingly, OCTA imaging revealed reduced vascular density, interrupted vascular continuity, and localized hypoperfusion within stenotic regions, indicating progressive microcirculatory impairment during fibrosis progression. In the C57 mouse bladder cancer model, tumor-bearing tissues displayed irregular surface morphology, heterogeneous scattering features, decreased vascular density, and disorganized vascular architecture, reflecting tumor-associated angiogenic remodeling and microvascular disruption.

Compared with conventional endoscopic imaging modalities, the developed endoscopic OCT/OCTA system provides simultaneous visualization of tissue microstructure and microvascular perfusion with micron-scale spatial resolution and without the need for exogenous contrast agents. The complementary NURD correction strategy significantly improved OCTA image stability and vascular continuity in rotational scanning environments, enhancing the applicability of proximally controlled endoscopic OCTA in deformable and dynamically moving luminal organs. These results demonstrate the capability of the system for reliable detection of disease-associated structural abnormalities and hemodynamic alterations in complex narrow-lumen environments.

In summary, this study developed and optimized a proximally controlled endoscopic OCT/OCTA imaging system for multi-organ in vivo imaging and pathological assessment in narrow-lumen organs. The system demonstrated robust adaptability to complex anatomical and dynamic imaging environments and enabled simultaneous visualization of tissue microstructure and microvascular perfusion across digestive, urinary, and vascular systems. Pathological imaging experiments further verified the capability of the system for detecting fibrosis-associated hypoperfusion and tumor-related vascular remodeling. These findings indicate that proximally controlled endoscopic OCT/OCTA possesses significant potential for early, minimally invasive diagnosis, dynamic disease monitoring, and intraoperative guidance in luminal organ diseases.

Organic light-emitting diode (OLED) is a new generation of display technology following LCD display, which has many advantages and has received widespread attention from researchers. OLED has been developing over thirty years, and emitters based on the three primary colors of red, green, and blue have met the requirements of commercialization. However, the latest Broadcast Television 2020 (BT.2020) display standard defines the color coordinates of blue emission as (0.131, 0.046), indicating higher requirements for the saturation of blue emission. Commercialized blue emitters are no longer able to meet the requirements of the BT.2020 display standard. Some studies have shown that color purity is not only related to the emission wavelength, but also to the full width at half maximum (FWHM) of the emission spectrum. Therefore, deep blue materials with narrow emission are the object of current research workers' pursuit, which is crucial for achieving ultra-high definition and full color with wide color gamut displays. Moreover, ultraviolet light, as the extension of deep blue toward the short-wavelength direction, is also widely concerned by researchers because of its scarcity. Therefore, developing high-efficiency near ultraviolet and ultraviolet luminescent materials with narrow half peak width is of great significance for academic research and commercial applications

All chemicals and reagents were purchased from commercial sources and used as received. The final product underwent vacuum sublimation to improve purity. All density functional theory (DFT) calculations were performed using the Gaussian 16 package. The theoretical calculation was completed at the M06-2X/6-31G (d,p) level. Solutions with a concentration of 1×10⁻⁵ mol/L were prepared for the solution measurements. All organic films used for the photoluminescence measurements were deposited onto clean quartz substrates via thermal evaporation at 1–1.5 Å s⁻¹ under high vacuum with a base pressure of < 10⁻⁵ torr (1 Å = 0.1 nm, 1 torr = 133.322 Pa). Ultraviolet-visible absorption spectra were measured on a Shimadzu UV-2600 spectrophotometer. PL spectra were recorded on a Horiba Fluoromax-4 spectrofluorometer. PLQYs were measured using a Hamamatsu absolute PL quantum yield spectrometer (C11347 Quantaurus_QY). Transient PL decay curves were measured using an Edinburgh Instrument FLS1000 spectrometer.

Cyclic voltammetry was conducted on a CHI 610E A14297 using a solution of tetra-n-butylammonium hexafluorophosphate (Bu₄NPF₆) (0.1 mol/L) in dichloromethane or dimethylformamide at a scan rate of 100 mV s⁻¹. A platinum wire was used as the auxiliary electrode, a glass carbon disk as the working electrode, and Ag/Ag⁺ as the reference electrode, with the redox couple ferricenium/ferrocene (Fc/Fc⁺) serving as the calibration standard. The ionization potential (IPCV) and electron affinities (EACV) of these molecules were calculated using the following formulas: IPCV = (E_ox − E_1/2(Fc/Fc⁺) + 4.8) eV and EACV = (E_red − E_1/2(Fc/Fc⁺) + 4.8) eV, where E_ox and E_red represent the onset oxidation potential and reduction potential relative to Fc/Fc⁺ (4.8 eV), respectively. Thermogravimetric analysis was performed on a Netzsch TG 209 under nitrogen flow at a heating rate of 10℃ min⁻¹. Differential scanning calorimetric was performed on a Netzsch DSC 200 F3 under nitrogen flow at a heating rate of 10 ℃ min⁻¹.

The glass substrates, precoated with a 90 nm layer of ITO with a sheet resistance of 15 to 20 ohms per square, were successively cleaned in ultrasonic bath of acetone, isopropanol, detergent, and deionized water, respectively, with each step lasting 10 min. Then, the substrates were completely dried in a 70 ℃ oven. To improve the hole injection ability of ITO, the substrates underwent O₂ plasma treatment for 6 min before fabrication. The vacuum-deposited OLEDs were fabricated under a pressure of < 5 × 10⁻⁴ Pa in the Suzhou Fangsheng FS-380 vacuum deposition system. Organic materials, LiF, and Al were deposited at rates of 0.5 to 1.5 A, 0.1, and 3 Å s⁻¹, respectively. The effective emitting area of the device was 9 mm², determined by the overlap between the anode and cathode. The luminance-voltage-current density characteristics and EL spectra were obtained via a PIN-25D silicon photodiode and Luminance Meter and an Ocean Optics USB 2000+ spectrometer, along with a Keithley 2400 Source Meter.

Based on the O/B/O multiple resonance molecular framework, benzocarbazole derivatives were introduced in the para position of the central benzene ring connected to B atom. By introducing methyl groups on the side of the intermediate benzene bridge, the emissions of the TB-PhCz, TB-MePhCz, and TB-2MePhCz were shifted from near ultraviolet to ultraviolet. Theoretical calculations indicate that a gradual decrease in the electron delocalization degree of S₁ states of TB-PhCz, TB-MePhCz, and TB-2MePhCz with an increase in the number of methyl groups introduced, exhibiting a characteristic of multiple resonance distribution. This is consistent with the gradual blue shift of their fluorescence emission peak positions, shifting from near ultraviolet to ultraviolet. Their doped devices all exhibit excellent electroluminescence performance, the maximum external quantum efficiency of TB-MePhCz based OLED device reaches 8.05% and thedevice has a high color purity (the full width at half maxima is 23 nm); TB-2MePhCz based OLED exhibits the deepest violet emission, thepeak of which is at 396 nm (the full width at half maxima is 20 nm), the CIE coordinates are (0.169, 0.012), and the CIE_y value approaches 0.01, which is currently the lowest CIE_y value reported in the literature, and the external quantum efficiency of the device reaches 7.81%, indicating that this molecular design strategy can effectively limit molecular conjugation and achieve blue shift of luminescent molecules, providing guidance for the development of more high-performance ultraviolet emitters.

Three ultraviolet-emitting molecules, TB-PhCz, TB-MePhCz, and TB-2MePhCz were designed and synthesized by tuning molecular torsion based on tert-butyl boron-oxygen multiple resonance skeletons. These molecules exhibit multiple resonance characteristics, achieving emission blue shift from near-ultraviolet to ultraviolet region. High-efficiency narrowband violet OLEDs were fabricated using the three molecules, respectively.

In fringe projection profilometry (FPP), the speed of fringe projection and image acquisition is a critical parameter determining the efficiency of three-dimensional measurement. Therefore, the rapid acquisition of fringe images has become an important research focus. Conventional methods based on DLP typically require a projector to occupy 255 pulse cycles for projection and rely on temporal integration imaging with a camera to obtain a single 8-bit sinusoidal fringe image. For phase-shifting (PS) algorithms that require at least three fringe images, this process is considerably time-consuming and fails to fully utilize the maximum refresh rate of the projector, making it difficult to meet the demands of dynamic 3D measurement. Moreover, the nonlinear responses of the projector and camera in traditional method introduce significant measurement errors, thereby reducing the overall measurement accuracy. To address these issues, this paper proposes a sinusoidal fringe image acquisition method based on line-shifted binary focusing projection.

The core of this paper is to propose a binary focused projection method based on line-shifted encoding to achieve efficient acquisition of sinusoidal fringe images. First, a line-shifted encoding strategy adapted to binary focused projection is proposed to realize the refined design of even-numbered binary encoded patterns. Within any period of the binary pattern, only one column is encoded as a normalized grayscale value of 1, and all other columns are assigned a value of 0. As the number of projection steps increases, different vertical patterns are projected one by one, and the encoded column set to a grayscale value of 1 switches column by column along the horizontal direction until it covers all horizontal columns in the projector's imaging plane. Second, a targeted sinusoidal fringe synthesis model is constructed to complete the accurate conversion of binary patterns into standard sinusoidal fringe images. This model is based on the principle of weighted grayscale superposition, assigns corresponding weights according to the contribution of each binary image in fringe synthesis, and synthesizes m binary encoded patterns into n standard sinusoidal fringe images with arbitrary steps through grayscale superposition operations. On this basis, a phase-shift algorithm is used to accurately extract phase information, and complementary gray codes are combined to assist in phase unwrapping, achieving high-precision and high-speed three-dimensional measurement.

To systematically verify the performance of the proposed method, three groups of controlled experiments are designed and conducted in this paper, with detailed experimental contents and results as follows: i) Performance verification experiment; it shows that as the coding period increases, the measurement accuracy gradually improves and tends to be stable when the period is 16 pixels. ii) Static comparison experiment; it shows that the measurement accuracy of the proposed method is significantly better than that of the binary defocus method, and slightly better than that of the traditional sinusoidal fringe projection method and the bitwise projection method. iii) Dynamic comparison experiment; it shows that the proposed method is suitable for dynamic measurement scenarios. Compared with the binary defocus method, although the proposed method needs to project more patterns, the measurement accuracy is significantly improved. Compared with the bitwise projection method, the proposed method can still achieve further improvement in accuracy under the condition that the number of fringe patterns is reduced by 63.64%. Compared with the traditional sinusoidal fringe projection method, although the number of required patterns is increased, since the digital projector can project binary patterns at a rate of thousands of frames per second, which is much higher than the hundred-frame level speed of projecting 8-bit sinusoidal fringes, the overall projection efficiency of the system is still effectively improved while ensuring the reconstruction accuracy.

In conclusion, this paper proposes a method for acquiring sinusoidal images based on line-shifted binary focused projection. This method performs focused projection on the encoded binary patterns and combines the constructed sinusoidal image synthesis model to synthesize multiple binary images into high-quality sinusoidal fringe images, thereby achieving the rapid acquisition of sinusoidal images. Verified by comprehensive comparative experiments under both static and dynamic scenarios, the proposed method can fully exploit the limit of the projector’s refresh rate and meet the application requirements of dynamic three-dimensional (3D) measurement. It not only significantly improves the 3D measurement speed, but also effectively suppresses the nonlinear errors of the measurement system, thereby further optimizing the measurement accuracy.

Pedestrian detection faces challenges from small-scale targets, dense object distributions, and complex backgrounds, which frequently result in insufficient detection reliability, missed detections, and false positives. Under such conditions, variations in pedestrian scale and scene complexity place higher demands on feature perception and multi-scale modeling. Existing detection models still exhibit limited robustness under small-scale and complex scene conditions, particularly when computational efficiency and model complexity are constrained, making it difficult to achieve a balance between accuracy and efficiency. These limitations remain critical obstacles to reliable pedestrian detection in practical scenarios. The objective of this study is to improve detection reliability and robustness under small-scale and complex scene conditions while reducing computational cost and model size, thereby enhancing inference efficiency and facilitating deployment in real-world pedestrian detection applications.

The proposed SRL-DETR is a lightweight pedestrian detection model developed based on the RT-DETR architecture. Firstly, a spatial-frequency feature fusion module was introduced for the backbone, integrating the SFSConv operator with a CSP structure. This design enhances the network’s perception of small-scale pedestrians by capturing both spatial and frequency features and facilitating more effective feature flow, enabling the model to extract richer and more discriminative information from low-resolution targets. Secondly, a refined feature scaling enhancement module was introduced to optimize multi-scale feature fusion in the neck network. SNI was introduced during upsampling to align high-level semantic features across scales, reducing information misalignment and preserving contextual consistency. GSConvE was introduced during downsampling to capture multi-scale textures and receptive fields, improving detection of dense pedestrians in complex scenes. These enhancements allow more accurate reconstruction and fusion of multi-scale features, strengthening the model’s robustness in challenging detection scenarios. Finally, LAMP was introduced to evaluate channel importance and remove less critical channels. This strategy effectively reduces model size and computation while preserving essential feature information, improving inference efficiency and making the model more suitable for deployment in resource-constrained environments. The proposed SRL-DETR model was evaluated on the CityPersons, TinyPerson, and WiderPerson datasets to validate both detection accuracy and generalization under varying pedestrian densities and scale distributions. The combination of enhanced small-target perception, refined multi-scale feature fusion, and adaptive pruning contributes to improved detection precision, robustness, and efficiency across different datasets.

The proposed SRL-DETR was evaluated on multiple pedestrian detection datasets, achieving high detection accuracy while simultaneously reducing model size and computational cost. On the CityPersons dataset, SRL-DETR achieved an mAP50 of 71.6%, representing a 2.0% improvement over the baseline model, while reducing parameters and computation by 50.5% and 42.1%, respectively. These results demonstrate that the proposed model maintains strong detection performance even under substantially reduced computational complexity. Comparison experiments further show that SRL-DETR outperformed YOLO and other compared methods across multiple evaluation metrics, indicating more balanced overall detection performance. In particular, SRL-DETR exhibited more stable detection behavior, suggesting improved capability in distinguishing pedestrians from complex backgrounds and reduced performance degradation caused by scale variation and scene clutter. On the TinyPerson dataset, SRL-DETR improved mAP50 by 3.7% compared with the baseline model. This improvement is mainly attributed to enhanced sensitivity to small-scale pedestrians and more effective multi-scale feature interaction, which alleviated missed detections and false positives in dense pedestrian scenarios. The results indicate that SRL-DETR is more effective in handling large scale variations and dense pedestrian distributions, which are commonly encountered challenges in pedestrian detection tasks.Similarly, on the WiderPerson dataset, SRL-DETR achieved a 1.0% improvement in mAP50 over the baseline, indicating consistent robustness across varying pedestrian densities and complex scene configurations. The stable performance across datasets with different data distributions further confirms the generalization ability of the proposed model under diverse detection conditions.Furthermore, the introduction of adaptive channel pruning enabled a substantial reduction in parameters and computation while preserving detection accuracy, highlighting the effectiveness of the lightweight optimization strategy. Overall, the coordinated improvements in small-scale target perception, multi-scale feature interaction, and model compression jointly contributed to improved detection accuracy, stronger robustness in dense and complex scenes, and enhanced efficiency, providing a practical and effective solution for pedestrian detection in real-world environments.

The proposed SRL-DETR model effectively balances detection accuracy, robustness, and computational efficiency for pedestrian detection under small-scale and complex scene conditions. Experimental results on the CityPersons, TinyPerson, and WiderPerson datasets demonstrate that SRL-DETR achieves consistent improvements in both accuracy and efficiency compared with the baseline and other compared methods, maintaining stable performance across varying pedestrian densities and scale distributions. The coordinated improvements enhance the model’s sensitivity to small-scale pedestrians, improve the handling of dense pedestrian distributions in complex scenes, and reduce model size and computation without compromising detection reliability. As a result, SRL-DETR maintains robust detection performance while satisfying efficiency requirements under practical constraints. In summary, SRL-DETR provides an effective pedestrian detection model that achieves a well-balanced trade-off between performance and computational efficiency, making it suitable for pedestrian detection tasks in small-scale and complex scene conditions.

High-precision component manufacturing imposes increasingly stringent requirements on surface defect inspection and three-dimensional (3D) topography measurement, particularly in advanced manufacturing fields such as semiconductor fabrication and precision micro–nano engineering. Laser confocal microscopy, owing to its excellent lateral resolution, strong optical sectioning capability, and robustness to environmental disturbances, has become an important tool for high-accuracy surface metrology. Nevertheless, conventional confocal microscopy systems are fundamentally constrained by their serial point-scanning imaging mode, which leads to limited imaging efficiency and restricts their applicability in high-throughput and real-time industrial inspection scenarios. To overcome these limitations, this study aims to develop a high-performance optical surface metrology system based on parallel-detection confocal microscopy, achieving a balanced optimization of imaging speed, resolution, and signal-to-noise ratio (SNR).

To address the throughput limitations of conventional point-scanning confocal systems, this study introduces a multi-detector parallel imaging architecture that enables simultaneous acquisition of multiple confocal signals, significantly improving data acquisition efficiency. The system employs a 405 nm laser as the illumination source and integrates a resonant-galvanometer scanning module to enable high-speed two-dimensional raster scanning. In the detection path, a fiber array comprising one central fiber and six peripherally arranged fibers couples the return beam into seven independent photomultiplier tubes (PMTs), facilitating parallel signal acquisition.

Based on this parallel detection framework, two complementary operational strategies are developed. First, a pixel-displacement rearrangement algorithm is implemented to fuse images acquired from multiple detectors. By integrating information from the central and peripheral detection channels, this approach effectively alleviates the inherent trade-off between SNR and spatial resolution encountered in traditional single-detector confocal systems, enabling high imaging quality to be maintained without sacrificing throughput. Second, by fully exploiting the spatial arrangement characteristics of the detector array, a dedicated fast scanning mode is developed. This scanning strategy allows multiple spatially offset detection channels to contribute to a single image frame, thereby enhancing temporal sampling efficiency.

To ensure stable and reliable system performance, several critical calibration and correction techniques are implemented. A multi-PMT gain calibration method is developed using a uniform illumination source to compensate for response variations among parallel detection channels, ensuring signal consistency and accuracy during multi-channel image fusion. For scanning distortion correction, a hardware-independent algorithm based on piecewise cubic Hermite interpolating polynomial (PCHIP) is proposed to address the nonlinear scanning distortion introduced by the resonant galvanometer. This method interpolates discrete distortion data obtained from standard grid imaging, maintaining high smoothness while exhibiting strong adaptability to complex and spatially varying distortion patterns. Additionally, to mitigate axial systematic errors and improve the reliability of 3D reconstruction, a software-based engineering axial calibration and compensation method is established. This approach uses a high-flatness reference mirror to acquire the system's inherent axial error distribution, which is then applied as a real-time correction factor for subsequent measurements.

Experimental evaluations validate the effectiveness of the proposed system and methods. In terms of imaging speed, the dedicated fast scanning mode achieves up to a 3.5-fold improvement in imaging frame rate compared with conventional confocal scanning approaches, while preserving spatial resolution and contrast. The pixel-displacement rearrangement algorithm enables the parallel detection framework to achieve a superior balance between SNR and resolution, outperforming conventional single-channel confocal imaging under identical gain settings.

For scanning distortion correction, quantitative evaluations demonstrate that the proposed PCHIP-based algorithm achieves a correction standard deviation of 4.151%, representing a 59.403% improvement in geometric correction accuracy compared with commonly used global quadratic fitting and local segmental fitting approaches. This significant enhancement confirms the method's superior adaptability to complex distortion profiles and its effectiveness in high-speed scanning scenarios.

In terms of axial measurement precision, the system retains the excellent optical sectioning capability of confocal microscopy. Experimental results demonstrate that the developed system achieves an axial measurement standard deviation of 4 nm when measuring a 1.8 μm standard height step using a 100×, 0.95 NA objective. The software-based axial calibration and compensation method effectively mitigates systematic errors, enhancing the consistency and robustness of 3D reconstruction results without increasing system complexity or requiring additional hardware.

The practical applicability of the system is further validated through two-dimensional and three-dimensional measurements of representative industrial samples, including semiconductor lithography structures and silicon-based wafer specimens. These experiments confirm that the system delivers excellent performance in both critical dimension detection and surface topography characterization, demonstrating its suitability for in-line inspection and quality control in advanced manufacturing environments.

This study successfully develops a high-performance optical surface metrology system based on parallel-detection confocal microscopy. By introducing a multi-detector parallel imaging architecture combined with a pixel-displacement rearrangement algorithm, the system effectively overcomes the inherent trade-off between SNR and spatial resolution that limits conventional confocal systems. The dedicated fast scanning mode, enabled by the spatial arrangement of the detector array, achieves up to a 3.5-fold improvement in imaging frame rate, delivering superior performance in imaging speed, resolution, and SNR. The complementary techniques developed—multi-PMT gain calibration, PCHIP-based scanning distortion correction (achieving 59.403% accuracy improvement), and software-based axial field compensation—provide a comprehensive solution for ensuring data fidelity, geometric accuracy, and measurement precision. With an axial measurement standard deviation of 4 nm on a 1.8 μm standard sample and demonstrated efficacy on semiconductor lithography and silicon-based wafer specimens, the system meets the demanding requirements for high-resolution, high-efficiency, and high-stability surface characterization in semiconductor manufacturing and precision engineering. Overall, this work provides an efficient, high-precision, and scalable optical surface metrology solution, offering significant potential for in-line inspection and quality control in advanced industrial manufacturing environments while establishing a feasible system architecture for developing advanced optical metrology instruments.

With the rapid development of aerospace engineering and the increasing demand for high-performance microsatellites, microwave plasma thrusters (MPT) have emerged as one of the most promising electric propulsion technologies for orbit maintenance, attitude control, station keeping, and deep-space exploration missions. Featuring high specific impulse, extremely low plume contamination, compact modular structure, and excellent long-term operational reliability, MPT offers significant advantages over traditional chemical propulsion systems, particularly in extending mission duration and reducing launch mass. However, the performance of MPT is fundamentally determined by the internal plasma characteristics, and the accurate, real-time, and non-intrusive diagnosis of key plasma parameters—most critically gas temperature—has long been a major technical bottleneck restricting further performance optimization and practical engineering application. Conventional diagnostic techniques such as Langmuir probes inevitably disturb the delicate plasma structure and suffer from severe probe tip ablation under high-temperature conditions, while laser-induced breakdown spectroscopy (LIBS) requires complex and expensive optical systems, making it impractical for continuous, full-domain monitoring inside the thruster resonant cavity. Therefore, developing a reliable, low-cost, and non-intrusive plasma parameter diagnostic method is of paramount importance for advancing MPT technology from laboratory research to industrial deployment.

Helium (He), an inert monatomic gas with a high ionization energy of 24.59 eV, abundant and well-separated characteristic spectral lines in the visible and near-infrared regions, and exceptional chemical stability, is selected as the ideal diagnostic propellant in this study. It effectively overcomes the inherent limitations of other commonly used working fluids: water vapor undergoes significant dissociation at high temperatures, producing complex molecular spectra that overlap with plasma emission lines; ammonia (NH₃) exhibits severe spectral interference due to its complex molecular structure; and hydrogen (H₂) has very few usable spectral lines within the detection range of most commercial spectrometers. In this work, a comprehensive non-intrusive parameter diagnosis system based on He atomic emission spectroscopy is established to achieve precise quantitative measurement of gas temperature. The experimental platform consists of a 2.45 GHz magnetron with continuously adjustable output power, a cylindrical resonant cavity designed to operate in the TM₀₁₁ mode, a high-precision mass flow controller, and a vacuum system maintained at approximately 3 Pa during all experiments. A NOVA-EX fiber-optic spectrometer equipped with thermoelectric cooling technology is employed to capture plasma emission spectra, with a detection range of 319–1106 nm, a slit width of 25 μm, a wavelength resolution of 0.76 nm, and an optical resolution of 1.97 nm. Five distinct He mass flow rates (100, 200, 300, 400, and 500 sccm) are tested at a constant microwave input power of 1100 W, and the Boltzmann plot method is applied to calculate gas temperature using four well-characterized He atomic transition lines (388.9 nm, 587.6 nm, 667.8 nm, and 706.5 nm). The Hα line at 656.3 nm, originating from minor air leakage into the cavity, is carefully excluded from the analysis. The local thermodynamic equilibrium (LTE) condition is rigorously validated by the high linearity of the Boltzmann plots, with all fitting R² values exceeding 0.85 and a maximum value of 0.974 achieved at 400 sccm.

Key experimental results demonstrate that the He mass flow rate exhibits a strong synchronous negative correlation with both the resonant cavity gas temperature and the thruster specific impulse. In the range of 100–400 sccm, both parameters decrease relatively gently due to stable microwave-plasma coupling efficiency and moderate energy transfer between excited and ground-state particles. However, when the flow rate exceeds 400 sccm, the decline rate increases dramatically. This abrupt change is attributed to three interconnected physical mechanisms: first, the specific energy (defined as the ratio of input microwave power to propellant mass flow rate) decreases significantly; second, excessive plasma density causes the plasma frequency to exceed the microwave frequency, leading to enhanced microwave shielding and skin effects that drastically reduce energy absorption efficiency; and third, increased particle collision frequency accelerates energy dissipation through inelastic collisions and thermal radiation. Under the LTE approximation, specific impulse is directly proportional to the square root of gas temperature, as temperature determines the average thermal kinetic energy of propellant particles and thus their exhaust velocity after expansion through the Laval nozzle. Furthermore, while variations in microwave input power do not alter the fundamental negative correlation between flow rate and performance parameters, they significantly affect the absolute values of temperature and specific impulse, as well as the critical threshold for abrupt performance degradation. Higher input power increases the specific energy per unit mass, shifting the critical threshold toward higher flow rates, whereas lower power reduces overall performance and shifts the threshold leftward. Extremely low power may even break the LTE condition, weakening the linear correlation in Boltzmann plots and reducing diagnostic accuracy. The measured gas temperature and specific impulse values show excellent agreement with previously published computational and experimental results, confirming the reliability and accuracy of the proposed diagnostic system.

This study provides a robust, non-intrusive method for high-precision parameter diagnosis of microwave plasma thrusters, reveals the intrinsic physical mechanisms governing the effects of He flow rate and microwave power on gas temperature and specific impulse, and clarifies the complex energy coupling processes between microwave fields and plasma. The findings offer valuable experimental data and theoretical support for propellant flow optimization, resonant cavity structural design improvement, microwave coupling efficiency enhancement, and numerical simulation validation of MPT. For practical engineering applications, low flow rates (100–300 sccm) are recommended for high-specific-impulse missions such as deep-space exploration, while the flow rate should be strictly controlled below 400 sccm to maintain stable and efficient propulsion performance. This work advances the development of accurate diagnostic technologies for microwave plasma thrusters and lays a solid foundation for their widespread engineering application in future aerospace propulsion systems.