Remote sensing images are formed by capturing the reflection or radiation information of electromagnetic waves from ground objects, reflecting the spatial distribution and semantic features of these objects. Remote sensing image target detection holds significant application value in fields such as environmental monitoring, military defense, and land resource management. However, due to the complex background of remote sensing images, significant variations in target scales, and a large number of small targets, existing target detection algorithms still exhibit deficiencies in feature extraction, multi-scale feature fusion, and model complexity control, making it difficult to simultaneously achieve both detection accuracy and computational efficiency. To address the aforementioned issues, this paper proposes an improved RTDETR remote sensing image target detection method based on an analysis of the structure and performance characteristics of the RTDETR algorithm. The aim is to enhance the model's ability to detect multi-scale targets and reduce model complexity while maintaining detection accuracy.

This article proposes an improved RTDETR remote sensing image object detection method to address the problems of insufficient feature extraction ability, limited multi-scale object modeling ability, and high model complexity in remote sensing image object detection tasks. Firstly, in terms of backbone network design, a lightweight backbone network StarNet is adopted to replace the original ResNet18. StarNet maps input features to a high-dimensional nonlinear feature space through element level multiplication, which can generate richer feature representations and enhance the model's ability to express complex remote sensing scenes. Secondly, to enhance the modeling ability of the model for targets of different scales, a deformable attention mechanism is introduced in the AIFI module of the efficient hybrid encoder to replace the original standard attention mechanism. The deformable attention mechanism combines the idea of sparse attention with dynamic shift strategy. By learning variable sampling positions, attention allocation becomes more flexible, allowing for adaptive attention to target regions of different spatial positions and scales. In addition, a bidirectional feature pyramid network module is introduced in the feature fusion stage of the efficient hybrid encoder. This module achieves full interaction and fusion of feature information at different levels through a bidirectional feature transfer mechanism from top to bottom and from bottom to top, effectively enhancing the model's ability to comprehensively utilize multi-scale features and further improving the detection performance of targets at different scales in remote sensing images. To verify the effectiveness of the proposed method, this paper conducted ablation experiments and comparative experiments on the DIOR remote sensing image dataset. At the same time, in order to evaluate the generalization ability of the algorithm in practical remote sensing application scenarios, further generalization analysis experiments were conducted on the RSOD dataset to reduce the risk of overfitting and verify the robustness of the model.

The experimental results show that the proposed improved algorithm has achieved significant performance improvement on the DIOR dataset. Among them, the accuracy and recall of the algorithm increased by 1.8% and 3.0% respectively, and the mAP50 and mAP50-95 indicators increased by 1.8% and 1.7% respectively. While improving detection performance, the parameter and computational complexity of the model were reduced by 37.7% and 32.5%, respectively, indicating that the proposed method significantly reduces model complexity while ensuring detection accuracy, achieving an effective balance between accuracy and efficiency. To further verify the generalization ability of the algorithm, this paper conducted generalization analysis experiments on the RSOD dataset. The experimental results showed that the improved algorithm improved the mAP50 and mAP50-95 metrics by 1.6% and 0.5%, respectively, while reducing the parameter count by 37.9%. The above results indicate that the method has good robustness and generalization performance on different remote sensing datasets, and can effectively alleviate overfitting problems. In addition, the detection performance of the algorithm proposed in this article was compared with various mainstream object detection algorithms on the DIOR dataset. The experimental results showed that the algorithm proposed in this article achieved the best performance in terms of average detection accuracy, showing significant advantages compared to other comparative algorithms. The main reason for the performance improvement lies in the synergistic effect of the introduced deformable attention mechanism and the weighted bidirectional feature pyramid network. The deformable attention mechanism can adaptively focus on key information regions at different scales through flexible attention allocation strategies, thereby enhancing the modeling ability of the model for local details and multi-scale features; The weighted bidirectional feature pyramid network achieves full integration of low-level detail features and high-level semantic features through a bidirectional information flow mechanism, effectively enhancing the ability to express multi-scale features. In terms of lightweight performance, the parameter and computational complexity of the improved algorithm are 12.4 M and 38.5 G, respectively, and the overall algorithm complexity is lower than other compared algorithms. This is mainly due to the adoption of the StarNet backbone network, which maps input features to a high-dimensional nonlinear feature space through element level multiplication in star operations. While generating rich feature representations, it significantly reduces model parameters and computational overhead, thereby enhancing the practicality and deployment value of the algorithm in remote sensing image object detection tasks.

This paper addresses the challenge of balancing accuracy and efficiency in remote sensing image object detection by improving the RTDETR algorithm and proposing a target detection method that considers both detection performance and model lightweighting. By introducing the StarNet backbone network, deformable attention mechanism, and bidirectional feature pyramid network, the model's ability to model and fuse features of complex scenes and multi-scale targets is enhanced. Experimental results show that this method achieves good detection performance on multiple remote sensing datasets and significantly reduces model complexity, validating its effectiveness and generalization ability. This research provides a reference for the deployment of remote sensing object detection algorithms in practical applications and resource-constrained platforms. In the future, we will further explore more efficient feature modeling methods to enhance the algorithm's applicability in complex remote sensing scenes.

Skin lesions are often accompanied by complex features such as blurred boundaries, variable shapes, and hair occlusion, which makes it difficult to accurately locate and finely characterize the lesion area in the segmentation task. At the same time, existing methods often have problems such as large number of parameters and high computational complexity, making them difficult to deploy on resource-constrained terminal devices. To this end, this paper proposes a novel lightweight skin lesion segmentation network based on dilated gating and adaptive bridging mechanisms.

The proposed method is designed to strike a balance between efficiency and segmentation performance, addressing the need for high-precision medical image analysis under resource-constrained environments.Specifically, a Dilated Gated Perceptual Enhancement (DGPE) module is introduced to improve both local detail modeling and global contextual understanding. This module employs a channel decoupling strategy combined with parallel depthwise separable convolutions of varying dilation rates to extract features at multiple scales. By integrating a gating attention mechanism and a progressive perception structure, the model effectively suppresses redundant background information and enhances edge structure preservation. To further bridge the semantic gap between encoder and decoder features, an Adaptive Feature Bridging (AFB) module is constructed. This module incorporates a dual attention mechanism—including Channel Attention Bridge (CAB) and Spatial Attention Bridge (SAB)—to perform cross-level feature calibration and establish global receptive field associations. In addition, a Cross-Granularity Feature Integration (CGFI) block is introduced to merge multi-resolution semantic features, enhancing the model’s capability to handle complex lesion structures with varying shapes and sizes. Furthermore, a Lightweight Initial Structure-Guided (LISG) module is developed to enhance shallow feature extraction. It integrates standard convolutions with large-kernel depthwise separable convolutions, forming long-range spatial dependencies in the early stages of the network. This is further supported by residual connections and cascaded small-kernel convolutions, ensuring effective gradient flow and feature stability across network layers. Lastly, an Efficient Channel Attention (ECA) mechanism is incorporated to model channel-wise dependencies using lightweight, adaptive one-dimensional convolution, effectively mitigating information loss during upsampling and preserving feature consistency throughout the decoding process.

Taking the ISIC2017 dataset as an example, our proposed method achieves mIoU and DSC metrics of 79.81% and 88.73%, respectively, representing performance improvements of 2.83% and 1.74% compared to the classic U-Net model. Simultaneously, the number of model parameters is significantly reduced by 25 times, and the computational complexity is reduced by 62 times, fully demonstrating its excellent computational efficiency while maintaining accuracy. Experimental results show that TransFuse and MHorUNet achieve mIoU of 79.21% and 79.37%, and DSC metrics of 88.40% and 87.99%, respectively, performing better than other methods compared in our paper. However, this is inseparable from their enormous computational burden. Compared to TransFuse, our proposed method reduces the number of parameters and computational complexity by 87 times and 51 times, respectively, significantly alleviating the hardware burden. Compared to MSS-UNet, another lightweight model, our proposed method reduces the number of parameters by 0.03M, a negligible difference, but achieves a 72-fold reduction in computational complexity. Even with the advancements in parameter compression achieved through dual-space shift MLP modules and external attention, MSS-UNet's computational complexity remains significantly higher than our proposed method, indicating room for improvement in computational efficiency. Furthermore, its high computational load may lead to overfitting, impacting model performance. In contrast, while UNeXt-S boasts a lower computational complexity of only 0.10 GFLOPs, its mIoU and DSC metrics are only 78.26% and 87.80%, respectively, showing a significant gap in segmentation accuracy. Its performance in other metrics is also inferior to our proposed algorithm. Our proposed method, in comparison, improves mIoU by 1.55% and DSC by 0.93%, demonstrating a superior balance between performance and efficiency.Comparative experiments show that our proposed method achieves an advanced balance between parameter count, computational complexity, and segmentation accuracy. Specifically, compared to traditional CNN-based medical segmentation models, our proposed method significantly reduces model redundancy and achieves more efficient feature extraction and fusion through structural design optimization. Compared to Transformer-based hybrid models, our method improves accuracy while significantly reducing computational resource consumption, thus alleviating deployment challenges. Furthermore, compared to similar lightweight models, our method not only ranks among the best in efficiency but also maintains a higher level of segmentation accuracy.

To address the common problems of insufficient segmentation accuracy and redundant model parameters in current skin lesion image segmentation methods, this paper proposes a lightweight skin lesion segmentation network based on dilatational gating and adaptive bridging. A dilatational gating perception enhancement module (DGPE) is designed, combining split dilatation units, gated attention units, and progressive perception enhancement units to fully mine multi-scale features and co-model global-local information, effectively mitigating the loss of feature details. An adaptive feature bridging module (AFB) is constructed, utilizing channel attention, spatial attention, and cross-granularity feature collaboration mechanisms to enhance effective interaction and fusion between cross-layer features, improving the accuracy of semantic information representation. A lightweight initial structure guidance module (LISG) is proposed in the initial stage of the network to improve the spatial representation ability and low-level semantic perception of shallow features, providing more robust guidance information for deep representation. Finally, efficient channel attention (ECA) is embedded in the upsampling stage, capturing the correlation features between key channels without increasing model complexity, and strengthening the localization ability of edge regions. This paper conducts a systematic evaluation on four publicly available datasets. Extensive experiments conducted on the ISIC2016, ISIC2017, and ISIC2018 datasets demonstrate the superior performance of the proposed model. It achieves mIoU scores of 85.78%, 79.81%, and 80.65%, and DSC scores of 91.90%, 88.73%, and 89.37%, respectively. With only 0.30 M parameters and 0.22 GFLOPs, the model significantly outperforms existing methods in terms of accuracy and efficiency, showcasing strong potential for real-world deployment in mobile medical applications and intelligent diagnostic systems, especially under constrained computing environments and limited storage conditions. Experimental results show that the proposed method significantly reduces the number of parameters and computational overhead while ensuring high segmentation accuracy. Its overall performance outperforms existing mainstream comparative models. Ablation experiments further validate the effectiveness and synergistic advantages of each module. This study provides a more reliable solution for the analysis of skin lesion images in clinical settings. Future work will explore integrating more feature enhancement strategies and structural optimization methods to further improve the model's generalization ability and make it applicable to research on other diseases.

In recent years, various devices for monitoring lower limb movement have emerged. Electrical sensors are susceptible to electromagnetic interference, while visual signals are affected by light intensity. Fiber Bragg Grating (FBG) sensors, on the other hand, have been widely adopted due to their excellent linearity, high sensitivity, flexibility, and strong anti-interference characteristics. During lower limb movement, the forces on the toes and heels of the foot continuously change, and the forces on the calf and thigh also fluctuate. However, research on the integrated monitoring of foot, calf, and thigh movements remains limited. Therefore, the development of a lower limb exoskeleton monitoring device is significant.

This study first established a theoretical model based on kinematics and verified its feasibility through theoretical analysis and simulation. Subsequently, a lower limb exoskeleton and its monitoring sensor unit were designed. Calibration experiments were conducted on the lower limb sensor unit using a tensile testing machine. The force sensitivities of all sensing units were obtained above 9.9 pm/N, and the goodness of fit values were all above 0.9906, demonstrating that the sensing units possess excellent monitoring performance. The exoskeleton provides support for the wearer, and the sensor unit integrated within it successfully monitors lower limb movements. After the volunteer wore the exoskeleton, motion monitoring experiments were conducted. The resulting data showed periodic changes, consistent with the conclusions drawn from the theoretical model and simulation analysis. The device features a simple power system that provides a certain level of assistance. Furthermore, the system includes an upper-level computer to analyze the monitored data and provide warning feedback based on the movement state.

In the theoretical model, as the angles of the three segments of the lower limb increase, the wavelength of the corresponding sensor units either increases or decreases. The relationship between step length and step height in the simulation analysis aligns with theoretical predictions. Additionally, dynamic analysis of the simplified structure further verified the accuracy of the step length-step height relationship. Finally, experimental data from the volunteer wearing the exoskeleton exhibited periodic changes, with stable mean values and standard deviations for the lower limb regions. The error between the theoretical and actual step length-step height was approximately 10%. The average value of the lower limb gait rhythm index (J) was 0.9379. These results confirm that the lower limb exoskeleton monitoring device meets monitoring requirements without negatively impacting the monitoring process and can accurately track the volunteer's lower limb movements. From a technical perspective, this study has made significant progress in sensing unit design and system integration. The advantages of FBG sensors lie not only in their high sensitivity and anti-interference capability but also in their suitability for multi-point synchronous monitoring due to their distributable embeddability and wavelength-encoding characteristics. In this system, by arranging sensing units at corresponding positions on the sole, ankle joint, knee joint, and thigh, a monitoring network covering key lower limb movement nodes was constructed. This distributed layout can capture the biomechanical changes during various phases of the gait cycle, providing multidimensional data for gait analysis and motor dysfunction assessment. In terms of system integration, the exoskeleton structure not only serves as a carrier for the sensing units but its mechanical design also fully considers ergonomics and movement adaptability. Lightweight hinge structures are used at the joints to reduce movement constraints while ensuring support stiffness. The power system employs lightweight electric actuators that primarily provide assistive torque during the gait propulsion phase, reducing the wearer's walking energy expenditure. When gait abnormalities (such as foot dragging or excessively high left-right asymmetry) or sudden force changes are detected, the system can provide alerts through sound or interface flashes, possessing preliminary gait anomaly warning capabilities.

Looking at the application prospects, this FBG-based lower limb exoskeleton monitoring device holds potential not only for motion analysis or assistance in healthy populations but also, and more importantly, in the field of rehabilitation medicine. It can provide objective and continuous gait parameter monitoring for patients after lower limb surgery or those with post-stroke sequelae, helping rehabilitation therapists quantitatively assess recovery progress. Furthermore, it could potentially integrate with functional electrical stimulation or robotic-assisted training to form an intelligent rehabilitation training system. Additionally, in specialized operational fields (such as firefighting or loaded marching), this technology could be used to monitor operators' fatigue states and prevent sports injuries. In summary, this study successfully designed and validated a lower limb exoskeleton monitoring device integrated with FBG sensing technology. The device achieves synchronous, anti-interference monitoring of multi-part lower limb movements with reliable data, while the exoskeleton structure provides a certain level of assistance during monitoring. The experimental results confirm their effectiveness, laying a solid foundation for subsequent research. Future work will focus on algorithm optimization, system lightweighting and portability, as well as clinical validation and functional expansion for specific application scenarios, aiming to ultimately transition the device from a laboratory prototype to a practical product.

With the rapid development of global air transportation system and unmanned aerial vehicle technologies, all kinds of aircraft are increasingly being applied for use in the security of civil and airspace monitoring and public safety. So, we need better and more correct aerial target recognition. Thanks to the development of deep learning and computer vision, Convolutional neural network based detection methods are being used for recognition of aircraft. Among them, the semantic segmentation can complete the localization of the pixel level target, therefore, it can provide much more detailed information to extract the aircraft contour and analyze the posture of the aircraft. However, compared to the medium and large aircraft, the small aircraft has fewer pixels in the image and its discriminative features are weaker, which is easily affected by the complex background and is hard to be precisely and steadily segmented by the current model. At the same time, in the practical application scenario, because of the difference in the use of imaging devices, the environment, and the distribution of data, it is easy to produce the domain shift phenomenon, and the model’s generalization ability is greatly reduced. And all these high-accuracy segmentation methods need very complex network structure and occupy lots of computational resources, which is hard to satisfy the real-time requirement of practical systems. So combining domain adaptation and light-weighting modeling approach to get both high accurate real-time small aircraft semantic segmentation of background. This is worth researching and putting into practical engineering.

The paper focuses on the problem of poor segmentation accuracy of small aircraft in cross-domain situation, and carries out some model design and exploration within the scope of a deep learning-based semantic segmentation framework. The classical U-Net plus an additional feature pyramid network (FPN) which is a multi-scale feature fusion structure to improve the model's ability to represent small aircraft targets. In addition to this, a Squeeze-and-Excitation(SE)channel attention is added to the network bottleneck such that it can learn to adjust the weight of each feature channel adaptively, thereby increasing the models ability to react to important discriminative features. To solve the different feature distribution caused by the different scenes and imaging conditions, we proposed a method to align the deep feature distribution between the source and target domains by the MMD, which can also enhance the model’s generalization to the target domain. On this basis, we use a teacher-student network and introduce a knowledge distillation approach, where the high-accuracy model with domain adaptation serves as the teacher. Using the teacher’s soft label to supervise the training of a light student model, we can greatly reduce the number of parameters and computational cost of the student model, but still have good segmentation effect. Experiment is conducted on the multi-scene small-aircraft dataset with explicit source-target domain split. To validate the effectiveness and real-time performance of the above proposed method through comparison experiment and ablation experiment respectively.

Small aircraft cross domain semantic segmentation performance, experimental results show that the proposed FPN-SE-DA-UNet is good at both accuracy and generalization. Feature pyramid network comes in to play in order to integrate information from different scale, and solve the loss of fine-grained features caused by the repeated down-sampling process. SE channel attention makes model better at noticing important discriminative features and leads to better target completeness and stability in segmentation results. Cross domain experiment shows that using the maximum mean discrepancy based domain adaptation to align the source and target domain feature can get much higher IOU and DICE score on the target domain test set, which is 0.5229 and 0.6866 respectively. Both is much better than all multiple semantic segmentation model. Also with the same model, it is possible to make a small student model by adding knowledge distillation. The student model is scored at 0.4557 IoU and 0.6260 Dice. The student model gets an average inference speed of 59.5 f/s on an NVIDIA GeForce RTX 3090, there is a slight drop in segmentation accuracy, this is about a 49.1% increase in real-time performance over the teacher model, it is a reasonable trade-off for accuracy vs. real-time performance. From the above we can see that the proposed method is able to obtain a stable small aircraft segmentation for cross-domain scenarios without sacrificing real-time application requirements.

To solve the problems of low segmentation accuracy, poor generalization ability, and weak real-time performance in cross-domain semantic segmentation of small aircraft, this study proposes a domain adaptation based semantic segmentation model, FPN-SE-DA-UNet, it introduces the FPN module to obtain richer multi-scale feature information, which can enhance the ability to capture small targets. On the other hand, the SE attention module is used to highlight more informative channel information in the network, and it introduces a domain adaptation strategy to align the features of the source domain and target domain, which can improve the model’s ability to generalize. On this basis, a lightweight student model is constructed with the help of a knowledge distillation strategy. It has good segmentation results and good inference efficiency, and the computational complexity has been greatly reduced. Experiments verify the effectiveness and practicability of the proposed method in cross domain semantic segmentation of small aircraft.

With the advancement of modernization and the development of urbanization, fire safety issues in various public venues, commercial buildings, and residential homes have become increasingly prominent. Fire incidents not only directly threaten lives and property but also generate smoke that often causes irreversible pollution and damage to the ecological environment. Traditional fire detection methods primarily rely on physical devices such as sensors, yet these approaches have limitations—including insufficient sensitivity, frequent false alarms and missed detections, susceptibility to environmental interference, and difficulty in achieving early warning. Consequently, there is an urgent need for more efficient fire detection algorithms. In recent years, rapid advancements in artificial intelligence have driven breakthroughs in deep learning within computer vision, particularly demonstrating exceptional performance in object detection. This has opened new avenues and research perspectives for fire detection studies.

Addressing current limitations in flame and smoke detection technologies—such as insufficient accuracy in complex scenarios, susceptibility to false negatives and positives, and high computational costs—this research innovatively applies the cutting-edge RT-DETR model from computer vision to fire detection tasks. The aim is to thoroughly explore this model's capability to capture and recognize flame and smoke features in complex, dynamic environments, striving to develop a fire detection algorithm with high detection accuracy, low false alarm rates, and the ability to meet real-time early warning requirements. RT-DETR is an efficient real-time end-to-end detector based on the Transformer architecture. It organically combines the local feature extraction advantages of convolutional neural networks (CNNs) with the global modeling capabilities of Transformers through the introduction of a hybrid encoder. By employing an innovative IoU-aware query selection mechanism to precisely filter initial target queries, it achieves a good balance between detection accuracy and inference speed while significantly reducing computational latency. However, when handling complex, multi-scale, and morphologically variable targets like flames and smoke, this model still suffers from insufficient capture of local information and limited extraction of fine-grained features. This often leads to false negatives and false positives for small targets, limiting its application in scenarios demanding high accuracy. To address this, this paper proposes a lightweight RT-DETR-based flame and smoke detection algorithm with multi-scale feature fusion to enhance the model's perception and discrimination capabilities for such targets. We first adopt RT-DETR-r18 as the baseline model, introducing the lightweight EfficientNetV1 architecture into the backbone network. Through evaluation and experimentation with mainstream lightweight networks like GhostNet, RepViT, and EfficientViT, results demonstrate that EfficientNetV1 achieves the optimal balance between accuracy and efficiency via its composite scaling strategy. It significantly reduces parameter count and computational burden while maintaining robust feature extraction capabilities, providing the optimal feature foundation for subsequent detection tasks. Second, the VoVGSCSP module is designed at the model's neck. This architecture integrates the efficient feature reuse of VoVNet with the gradient splitting concept of CSPNet, effectively enhancing the model's expression capabilities for key features like flame-smoke texture and color while suppressing feature redundancy. Finally, addressing the core challenge of extreme scale variations in flame-smoke targets, this paper introduces the more efficient multi-scale feature fusion mechanism bidirectional feature pyramid network (BiFPN)-E. During the weighted fusion process in the BiFPN, it innovatively employs the ELU activation function for weight normalization. The non-zero saturation property of the ELU function when handling negative inputs generates smoother and more discriminative fusion weights, thereby strengthening the interaction and complementarity between feature maps at different scales. This significantly improves the model's perception and discrimination accuracy in complex scenarios such as large-scale flames, faint smoke, and occluded objects.

The final experimental results fully validate the comprehensive effectiveness of the proposed method. On the NEWFire dataset, the improved model achieved mAP50 and mAP50-95 scores of 84.5% and 57.0%, respectively, representing improvements of 1.1% and 1.7% over the baseline RT-DETR model. On the self-built E-fire dataset, the accuracy metrics mAP50 and mAP50-95 improved by 1.8% and 2.2%, respectively. More importantly, while achieving these accuracy gains, the model demonstrated significant lightweighting: model parameters were reduced by 32%, and floating-point operations were drastically reduced by 65%. In-depth ablation experiments further confirm that the BiFPN-E architecture is the most critical contributor to accuracy gains, while the EfficientNetV1 backbone serves as the core for achieving efficient model compression. Visual analysis also reveals that the improved model's feature heatmap responses are more concentrated on actual flame-smoke regions, demonstrating stronger suppression of fire-like distractors and complex backgrounds, higher recognition accuracy, and effectively reducing both false negatives and false positives.

Through systematic enhancements to the RT-DETR model, this research achieves a synergistic breakthrough in accuracy, efficiency, and feature discriminative power for fire detection. The EfficientNetV1 backbone enables efficient model compression, while the introduced VoVGSCSP module enhances the model's ability to capture irregular fire shapes and dynamic changes. The BiFPN-E architecture significantly improves detection accuracy through optimized multi-scale feature fusion. The synergistic effect of these three components enables the model to maintain high accuracy while reducing parameters by 32% and computational load by 65%. This breakthrough overcomes computational resource constraints while enhancing performance, providing a practical solution for achieving high-precision real-time fire alerts on resource-constrained edge devices.

Precision cutting tools, as the core components for the construction of national key equipment, have attracted growing attention to their quality. Micro-scale defects on the tool surface, such as cracks and bumps, expand rapidly under high-intensity machining environments. The defects at the damaged parts will be continuously amplified, which affects the performance and safety of end products, also severely restricts the efficiency of product processing and manufacturing. The defects on the surface of cutting tools exhibit much smaller morphological and texture features, and this greatly increases the difficulty of automated detection. Therefore, the development of efficient and accurate micro-scale defect detection technology for tool surfaces has become a key factor in intelligent manufacturing. Although a large number of research methods related to different attention mechanisms have been proposed in the published literature. However, when these methods are directly applied to detect tiny appearance defects on the tool surface, the following two key problems will arise: 1) the problem of difficulty in identifying micro-scale defects. When existing attention mechanisms deal with the morphological diversity of tiny defects on the surface of precision cutting tools, they lack effective coupling and expression of channel features, resulting in the inability to accurately identify key information. 2) The problem of defect texture diversity. When facing the diverse texture features of complex defects on the surface of precision cutting tools, traditional methods fail to make full use of frequency domain information to capture subtle changes in defects, which affects the accuracy of detection.

To address these issues, a precision tool defect detection model is proposed based on adaptive multi-parallel attention (AMPA). The core architecture of AMPA consists of two functionally complementary sub-modules: the star channel attention module (SCAM) and the frequency-spatial collaborative attention module (FSAM). Adopting a parallel processing mechanism, the two modules mine defect features from different dimensions: SCAM focuses on the channel dimension, strengthens the nonlinear coupling between channels through star operations, and captures the key channel responses of micro-scale defects; FSAM integrates the frequency and spatial dimensions, extracts full-frequency domain information using fast Fourier transform (FFT), and fuses spatial structural features to address defect localization under complex textured backgrounds. This parallel design can maximize the preservation of the uniqueness of the two types of features-SCAM and FSAM have low feature mutual information (low redundancy), and the total information entropy is higher than that of serial structures (preventing subsequent modules from overwriting preceding features), laying a foundation for efficient feature fusion. From the aspects of actual processing lists as follows. Initially, a star-channel attention module (SCAM) is constructed, utilizing star pperations to enhance nonlinear coupling between channels, thereby improving the representation ability of key channel features. Secondly, a frequency-spatial attention module (FSAM) is designed to combine frequency domain and spatial information, effectively capturing low-frequency, mid-frequency and high-frequency features as well as spatial structural information in images. Finally, the fusion weights of these two modules are dynamically adjusted through learnable parameters to achieve adaptive feature fusion, enhancing the detection performance.

The dual improvements of the proposed methods demonstrate that while incurring a controlled computational cost, the AMPA module not only enhances detection accuracy under a single threshold, but also strengthens the model’s comprehensive detection capability across multiple scenarios and evaluation criteria, fully embodying its application value in multi-category industrial defect detection. Experimental results demonstrate that integrating the proposed method into mainstream YOLOv11 model can achieve a mAP0.5 of 90.2% on a self-constructed precision tool defect dataset, achieving an improvement of 2.4% compared with the traditional YOLOv11. This result validates the model's effectiveness in detecting micro-scale target defects on tool surfaces. Additionally, on the publicly available NEU steel surface defect dataset, the proposed method can reach a mAP0.5 of 76.1%, surpassing traditional YOLOv11 by 1.3%, demonstrating the model's robust generalization capability. In summary, by fusing channel, frequency-domain and spatial features, the AMPA module not only improves detection accuracy under a single threshold, but also boosts robustness across different IoU criteria, effectively optimizing the detection performance for multi-category steel surface defects and verifying the effectiveness and practicality of the proposed method.

Aiming at the problems of existing attention mechanisms in the detection of micro-defects on the surface of precision cutting tools, an adaptive multi-parallel attention (AMPA) module is proposed in this paper. First, the star-channel attention module (SCAM) and the frequency-spatial attention module (FSAM) are executed in parallel, combined with learnable dynamic weights, to realize the adaptive fusion of channel-dimensional features and frequency-spatial features. To verify the effectiveness of the proposed method, the AMPA module is embedded into mainstream YOLO series models (from YOLOv8 to YOLOv11) respectively, and evaluated on a self-built precision cutting tool dataset, achieving a maximum improvement of 2.7% in detection accuracy. For testing the generalization ability of the algorithm, further experiments are conducted on the NEU steel surface defect dataset, with a maximum accuracy improvement of 1.5% obtained. Experimental results show that the AMPA module not only significantly optimizes the detection performance of small-target defects on precision cutting tools, but also effectively enhances the recognition performance of multi-category complex defects on steel surfaces, demonstrating its wide applicability in enhancing feature representation and improving model performance across different industrial scenarios. Although the proposed method shows effectiveness in the detection of micro-scale single-category tool defects, it still has limitations in complex industrial scenarios: its adaptability to multi-category micro-scale defects needs to be enhanced, its robustness against interferences such as extreme illumination and blurring should be improved, and the real-time performance under lightweight deployment still has room for optimization. Future research will further improve the method by introducing lightweight module designs and other strategies.

Steel surface defects such as cracks, scratches, rolled-in scale, pitting, patches, and inclusions are common in industrial manufacturing and may cause severe degradation in corrosion resistance, wear resistance, and mechanical reliability, which makes accurate and efficient surface defect inspection essential for quality control. Although deep learning based detectors have achieved promising results, reliable detection of tiny defects remains difficult because these defects often present small size, low contrast, blurred boundaries, and large scale variations, while industrial backgrounds contain complex textures and noise that easily interfere with feature extraction. Two-stage detectors usually require heavy computation and fail to meet real-time requirements, and many one-stage detectors still suffer from high miss rates on small defects and limited cross-scale adaptability. RT-DETR introduces an end-to-end real-time detection framework by eliminating non-maximum suppression and leveraging Transformer-based global modeling, which improves dense target prediction, yet its parameter redundancy and insufficient cross-resolution feature alignment still restrict its efficiency and robustness in industrial scenarios.

To address these limitations, this study proposes RMD-DETR, an improved RT-DETR based detection algorithm that enhances detection accuracy, cross-scale stability, and inference efficiency through three coordinated modifications. First, a lightweight FasterNet backbone is introduced to replace the original feature extractor, and the partial convolution in FasterNetBlock is further optimized with a reparameterization strategy to strengthen feature representation under limited computation. During training, multi-branch convolutional paths are used to enrich spatial diversity and improve learning capacity, while during inference these branches are equivalently fused into a single convolution operation, which simplifies the computational graph and accelerates deployment without sacrificing accuracy. Second, a dynamic multi-scale feature fusion network (DMSFPN) is designed as the neck to alleviate semantic inconsistency and misalignment between high-level semantic features and low-level detailed features, which are critical for localizing micro-defects. DMSFPN integrates a multi-scale convolution block (MSCB) to capture context information at different receptive fields, where heterogeneous kernel branches focus on tiny, medium, and large defect patterns, and a channel shuffle mechanism is employed to enhance cross-channel interaction and reduce feature redundancy. To improve detail recovery in cross-scale aggregation, an efficient upsampling convolution block (EUCB) is introduced, which combines lightweight convolution operations with deformable convolution based adaptive sampling, thereby mitigating the feature blurring problem caused by standard interpolation and preserving spatial continuity for small targets. A bidirectional feature fusion strategy is adopted to enable both top-down semantic propagation and bottom-up detail enhancement, improving cross-resolution consistency while avoiding redundant repeated computation. Third, to enhance feature interaction under cluttered backgrounds, a token statistics self-attention (TSSA) mechanism based on statistical feature compression and variational rate reduction is incorporated into the adaptive feature interaction module, forming an AIFI-TSSA enhancement block. Unlike conventional self-attention that computes pairwise token similarity with quadratic complexity and high memory cost, TSSA performs global context modeling with linear complexity through token statistics driven projection and denoising, which reduces computation overhead and improves robustness against redundant textures, occlusion, and noise. The proposed RMD-DETR is evaluated on three industrial datasets, including NEU-DET, the Tianchi Aluminum profile surface defect dataset, and GC10-DET, and comprehensive experiments are conducted to verify the effectiveness of each proposed component, including ablation studies, comparative experiments with representative detectors, cross-scale evaluation by resizing inputs to a smaller scale, and generalization tests across different industrial scenes.

Experimental results on NEU-DET demonstrate that RMD-DETR achieves 76.7% mAP@0.5, outperforming the baseline RT-DETR by 2.6%, while precision and recall reach 77.8% and 72.6%, indicating improved localization reliability and reduced missed detections. The model simultaneously reduces parameters and computational cost by 39.9% and 39.3%, and the inference speed increases to 78.9 f/s, satisfying the real-time inspection requirement of industrial production lines. In cross-scale detection experiments, the missed detection rate decreases from 55.4% to 50.9%, showing stronger adaptability under significant scale variation. Generalization experiments further confirm the transferability of the proposed design, where mAP@0.5 improves by 2% on the Tianchi dataset and by 3% on GC10-DET compared with the baseline model. Visualization analysis under Gaussian noise demonstrates that the proposed method maintains higher confidence and more complete defect coverage, especially for micro-scratches and fine cracks that are easily ignored in complex textures. These results indicate that the three improvements are complementary: the reparameterized lightweight backbone improves efficiency while retaining feature extraction capability, DMSFPN enhances multi-scale semantic alignment and detail preservation, and AIFI-TSSA provides stable global modeling with reduced complexity, enabling the detector to focus on discriminative defect regions rather than background patterns.

In conclusion, RMD-DETR provides an effective balance between accuracy and speed for steel surface defect detection, achieving enhanced robustness and cross-scale performance across multiple industrial datasets, and it is suitable for practical quality inspection systems requiring high throughput and reliable detection. Future work will explore further compression for extremely resource-constrained edge devices and deployment optimization through quantization and acceleration tools to improve engineering applicability.

Sleep monitoring is critical for early disease warning and personalized health management as respiratory rate, heart rate and body movement during sleep directly reflect physiological and pathological states. Traditional monitoring methods, such as electrocardiographs (ECGs) and polysomnographs (PSGs), are invasive, requiring electrode patches to be attached to the human body. Despite their high precision, they are poor in comfort, and tend to disrupt natural sleep. In contrast, non-invasive optical fiber sensors possess advantages including small size, electromagnetic interference (EMI) resistance, and high sensitivity, which enable them to perceive physical quantities such as vibration and pressure via wavelength or light intensity modulation, thus exhibiting significant potential in sleep monitoring. Nevertheless, noise interference persists during the processes of photoelectric conversion and signal acquisition, which impairs the detection accuracy of respiratory and cardiac signals. Existing denoising algorithms like wavelet soft thresholding have constant amplitude deviations, and offline-dependent models fail to meet real-time needs. This study aims to develop an all-fiber multimodal sleep monitoring system for non-invasive, high-precision and real-time detection of sleep respiratory rate, heart rate and turnover state, providing a reliable technical solution for home sleep health monitoring.

A sleep monitoring system based on an all-fiber multi-modal sensing network was constructed. A 40 cm×30 cm sandwich structured flexible monitoring pad was fabricated, embedded with macrobend 1060XP single-mode fiber and dual fiber Bragg gratings. The macrobend fiber was optimized to a 20 mm bending radius with nine loops through experiments, arranged in two rows of semi-circular curves, embedded in a latex pad's middle layer with silicone sheet covers to enhance cardiopulmonary vibration sensitivity and structural durability. Dual FBGs with initial wavelengths of 1548.625 nm and 1548.473 nm are symmetrically spaced 20 cm apart, bonded to a rigid PVC substrate using epoxy AB glue and fixed on the pad to offset temperature drift and capture turnover-induced strain. For signal processing, an improved threshold λ was proposed to achieve adaptive thresholding across different decomposition layers, considering the impacts of decomposition levels j and high-frequency wavelet coefficients cdn on threshold selection. Compensation term decay rate parameter α and exponential term growth rate β were introduced to optimize the threshold function. These two parameters dynamically regulate the output value of the threshold function when λ is applied, ensuring continuous and differentiable processing of cardiopulmonary signals while balancing noise suppression and weak signal preservation. Subsequently, second-order Butterworth band-pass filters are configured for respiratory and heartbeat signals, with the respiratory band 0.133–0.417 Hz and the heartbeat band 0.833–1.67 Hz to separate mixed signals, and fast Fourier Transform (FFT) is used to extract characteristic frequencies. After that, a body movement determination threshold was constructed based on the wavelength difference change rate of dual FBGs to distinguish left and right turns and count turn events. Experimental validation was conducted in a constant-temperature laboratory with a 61 kg subject lying supine on the monitoring pad and E-HA03 commercial electrocardiograph data was served as reference.

Practical tests showed the improved denoising algorithm achieved a signal-to-noise ratio of 24.02 dB and a root mean square error of 0.09, outperforming traditional hard threshold method with signal-to-noise ratio of 14.32 dB and root mean square error of 0.27 as well as soft threshold method with signal-to-noise ratio of 7.12 dB and root mean square error of 0.62. Respiratory spectrum peaked at 0.33 Hz corresponding to 20 breaths per minute, heartbeat at 0.97 Hz, corresponding to 58 beats per minute. Compared with E-HA03 data 19 breaths per minute and 61 beats per minute, Respiratory rate error was 1 breath per minute and heart rate error was 3 bpm, meeting clinical home monitoring requirements. A turnover determination model was established based on the wavelength difference change rate of dual FBGs, where the left turnover threshold was set to ≥2.8 nm and the right turnover threshold to ≤−2.4 nm for distinguishing turnover directions and counting events. During the test, three left and three right turnovers were detected at 493, 530, 568, 597, 791, and 820 s, with no false or missed detections despite minor limb movements. The dual FBG design effectively offset temperature drift, with approximately stable wavelength baselines after thermal equilibrium. These results confirm the algorithm's robustness and reliability in sleep scenarios.

The non-invasive flexible pad ensured user comfort without disrupting sleep, and three key advancements are achieved: 1) the integration of macrobend fibers (optimized for bending radius and layout) and symmetrically arranged dual FBGs breaks the constraints of single-sensor systems, where the fibers capture subtle cardiopulmonary vibrations and the FBGs realize precise body-turnover detection to enable synchronous multimodal sensing of cardiopulmonary signals and body-turnover states. 2) The dual-parameter adaptive wavelet threshold denoising algorithm balances noise suppression and weak signal preservation, successfully capturing faint heartbeat signals that are easily masked by respiratory signals. 3) Symmetric dual FBGs combined with wavelength difference analysis offset temperature drift, enhancing the system's stability in complex home environments. The all-fiber multimodal sleep monitoring system achieves high-precision non-invasive monitoring of respiratory rate, heart rate and turnover state. Its relatively excellent denoising performance, efficient signal separation and accurate turnover recognition provide a new technical approach for home sleep health monitoring, with broad application prospects in elderly care and personalized health management. Future work will optimize sensor array density to distinguish more turnover-related movements such as limb movements and expand clinical trials to diverse populations to enhance generalizability.

Polarization control in the near-infrared (NIR) regime serves as a fundamental technology for numerous advanced applications, including high-capacity optical communication, high-resolution biomedical imaging, molecular spectroscopy, and optical sensing. Traditional tools for polarization manipulation, such as wave plates, prisms, typically exhibit significant limitations in terms of size, bandwidth, functional versatility, and integration potential with modern photonic circuits. The advent of optical metasurfaces has revolutionized the mentioned limitations, enabling precise control over the amplitude, phase, and polarization of light within a ultra-thin planar film. However, a persistent challenge in this field is the realization of passive metasurfaces that can simultaneously provide multiple polarization transformation functions with high conversion efficiency and ultra-broad bandwidth. Most reported designs are restricted to a single operating mode or rely on external active tuning mechanisms, which substantially increase complexity of the system.

To directly address the technological challenge, in this paper, we introduce a comprehensive design, rigorous numerical analysis, and an elucidation of the operating principle of a novel, fully passive reflective metasurface designed for ultra-broadband, high-efficiency multifunctional polarization conversion across multiple bands within NIR. The unit cell is meticulously constructed from three distinct layers: a top-layer nanostructured resonator made of silver, a middle dielectric spacer made of silica, and the bottom layer of silver film. The key innovation lies in the geometry of the top-layer resonator, which is an anisotropic notched rectangular ring — a design deliberately lacking fourfold rotational symmetry to break the symmetry for differently polarized incident light.

Finite element method (FEM) is employed for all simulations. Simulation boundaries employed periodic conditions in the x-y plane and perfectly matched layers in the propagation direction. Performance evaluation is conducted under normal incidence of x polarized plane wave. A comprehensive set of metrics is calculated to quantify the performances, including co-polarized reflection coefficient, cross-polarized reflection coefficient, polarization conversion ratio, reflectance of the converted circular polarization states, as well as derived polarization rotation angle and ellipticity angle. To reveal the underly fundamental physical principles, an anisotropic decomposition method is employed to analyze the amplitude and phase responses of the electromagnetic fields along the two principal axes of the resonator. Additionally, surface current distributions and electric field distributions are examined in detail at multiple wavelengths to visualize the excitation of electric dipole moments.

Simulation results validate the exceptional multifunctional polarization conversion capabilities and broadband characteristics of the proposed metasurface structure. First, it operates as a highly efficient half-wave plate within the wavelength range from 1520 nm to 1720 nm. In this range, the cross-polarized reflectance consistently exceeds 0.8, reaching a peak of 0.939 at the wavelength of 1620 nm, while the undesired co-polarized reflection is suppressed to near zero (0.0009 at the same wavelength). The corresponding polarization conversion ratio remains above 0.9 within the wavelength range of 1570−1735 nm, reaching a near-ideal maximum of 0.999 at 1650 nm, while the polarization rotation angle is maintained at approximately 90° over a broad wavelength range from 1380 nm to 1880 nm, enabling orthogonal polarization conversion. Second, the metasurface seamlessly transitions to broadband quarter-wave plate in the two adjacent wavelength bands. In the lower wavelength band from 910 nm to 1520 nm, it predominantly converts x-polarized incident light into right-handed circularly polarized light, with a reflectance rxR greater than 0.8 and a peak efficiency of 0.965 at 1350 nm. In the higher wavelength band from 1720 nm to 2500 nm, the x-polarized incident light is converted to left-handed circularly polarized light, with a reflectance rxL > 0.8 with a peak reflectance of 0.958 at 1910 nm. The ellipticity angle η quantitatively determines the degree of circular polarization, confirming the polarization conversion by remaining around −45° in the band generating right-hand circular polarization (1080−1440 nm) and around +45° in the band generating left-hand circular polarization (1800−2500 nm). The physical origin of the behavior can be traced back through anisotropic decomposition analysis, which reveals that the critical phase difference between the u and v field components is approximately −3π in the wavelength band working as half-wave plate, satisfying the condition for a 180° phase shift. In the wavelength band working as quarter-wave plate, Δφ keeps unchanged at around −3.5π and −2.5π, respectively, meeting the requirements for orthogonal polarization conversion. Surface current analysis at representative wavelengths (1130 nm, 1620 nm and 2180 nm) indicates the excitation of strongly directional electric dipoles on both the top-layer resonator and the bottom substrate. These dipoles excite electromagnetic coupling, facilitating the phase modulation required for efficient and broadband conversion.

In conclusion, in this work, we successfully demonstrate the design and theoretical validation of a passive reflective type metasurface that integrates high-efficiency half-wave plate and quarter-wave plate functionalities within a single, ultra-thin device in the near-infrared. By exploiting the geometric anisotropy, the proposed design overcomes the traditional trade-offs between bandwidth, efficiency, and functional versatility. Through detailed numerical investigations, together with analyses of polarization ellipse parameters and current distribution, a thorough understanding of the underlying conversion mechanisms is provided. It holds immense promise for revolutionizing compact photonic systems, with potential applications including polarization-encoded optical communications, polarization-sensitive endoscopic imaging, miniaturized spectroscopic sensors, quantum optical devices, and next-generation lidar systems, offering a new design approach for multifunctional polarization control.

Unmanned aerial vehicle (UAV) imagery poses enduring challenges for object detection, largely because targets are frequently small, densely distributed, and captured under heterogeneous imaging conditions. In contrast to ground-level scenes, aerial views often exhibit pronounced scale variation, viewpoint changes, partial occlusions, and severe background clutter arising from man-made structures, vegetation, road textures, and illumination-induced shadows. In addition, motion blur and sensor noise are common due to platform dynamics and altitude-dependent resolution constraints. Collectively, these factors exacerbate the loss of fine-grained cues during feature downsampling and increase both missed detections and false alarms, particularly when small objects occupy only a few pixels and their boundaries are confounded with background patterns.

This work presents an attention-fused object detection approach built upon an improved YOLOv8n framework to alleviate the above limitations. Although YOLOv8n is attractive for edge deployment owing to its lightweight design, its baseline configuration can underperform on UAV datasets when high-frequency details are attenuated by successive convolution and pooling operations and when multi-scale feature aggregation is insufficiently discriminative. To strengthen representation learning while retaining computational efficiency, we introduce three complementary modifications: (i) a backbone replacement based on FasterNet to better preserve small-object structure, (ii) an inter-stage Triplet attention mechanism to model coupled channel-spatial dependencies, and (iii) a neck-level enhancement module that integrates receptive-field attention convolution (RFAConv) with the convolutional block attention module (CBAM) via a newly designed DARFCB-C2f block.

First, the original YOLOv8n backbone is replaced with FasterNet to improve the retention of fine details that are critical for detecting small targets. FasterNet employs partial convolution (PConv) in conjunction with pointwise convolution (PWConv) to balance representational capacity and computational cost. By applying convolution to only a subset of channels, PConv reduces redundant computation while maintaining local structural information, whereas PWConv performs channel mixing to sustain expressive feature interactions. Notably, the combination of PConv and PWConv yields a T-shaped restricted receptive-field pattern that biases attention toward central feature-map regions. Given that small objects typically occupy limited spatial extents, preserving and emphasizing central responses can mitigate the information loss induced by downsampling. As a result, the revised backbone enhances small-object feature fidelity without substantially increasing model complexity.

Second, to better accommodate the coexistence of multiple target categories and large scale disparities within a single aerial frame, Triplet attention is inserted between the backbone and the neck. Conventional attention modules often compute channel and spatial attention independently, which may limit their ability to capture joint interactions across dimensions. Triplet attention addresses this by modeling dependencies through three coordinated branches: channel-spatial height, channel-spatial width, and spatial height-width. This triadic formulation enables richer coupling between channel and spatial representations, thereby providing a more comprehensive description of salient object patterns across scales. In dense UAV scenes, where small and medium-sized objects may appear simultaneously and compete for representational resources, the proposed attention insertion improves feature selectivity and enhances robustness to scale variation.

Third, we redesign the neck feature-fusion blocks by proposing a DARFCB structure and constructing a DARFCB-C2f module that replaces the standard C2f blocks in YOLOv8n. The design leverages RFAConv to expand the effective receptive field and to perform attention-guided feature fusion, which is beneficial for incorporating contextual cues without sacrificing boundary-sensitive information. Such boundary cues are particularly important in UAV imagery, where object contours are frequently degraded by blur, low resolution, and background interference. In addition, CBAM is integrated to refine intermediate representations via channel and spatial attention, amplifying informative regions while suppressing irrelevant responses. Together, RFAConv and CBAM reinforce edge-related features and improve the reliability of multi-scale aggregation, leading to more stable detection under complex environmental conditions.

Comprehensive experiments are conducted on the VisDrone2019 benchmark, which is widely used for UAV object detection and contains challenging scenarios including dense traffic, crowded pedestrian regions, substantial scale variation, and cluttered backgrounds. Ablation studies and module replacement analyses demonstrate that each proposed component contributes measurable gains, and their combination yields the strongest overall performance. Relative to the baseline YOLOv8n detector, the proposed method improves precision (P) by 4.9%, recall (R) by 3.9%, and mAP50 by 6.1%. These improvements reflect a meaningful reduction in both false positives and false negatives, indicating that the attention-fused design effectively enhances the discriminability and completeness of detection results in UAV scenes.

To assess cross-domain robustness, additional evaluations are performed on the WiderPerson and PASCAL VOC datasets, which differ from VisDrone2019 in terms of scene composition, object density, and imaging characteristics. The proposed model maintains strong performance across these datasets, suggesting that the introduced backbone redesign, attention fusion strategy, and neck enhancement provide transferable benefits rather than dataset-specific overfitting. This generalization evidence supports the applicability of the approach to practical UAV-driven tasks such as aerial surveillance, traffic monitoring, and search-and-rescue operations.

In summary, this paper proposes an attention-fused YOLOv8n-based detector for UAV aerial imagery that integrates three synergistic improvements: a FasterNet backbone to preserve small-target details and alleviate downsampling-induced information loss, triplet attention to capture richer multi-dimensional dependencies for multi-scale recognition, and a DARFCB-C2f neck module that combines RFAConv and CBAM to strengthen edge features and improve robustness in cluttered environments. Experimental results on VisDrone2019 and cross-dataset validations on WiderPerson and PASCAL VOC collectively confirm that the proposed modifications substantially improve detection accuracy while retaining the efficiency advantages of a lightweight one-stage detector.

Light is the most fundamental medium through which humans perceive the objective world, encompassing multiple physical dimensions such as amplitude, phase, and polarization, and these dimensions carry abundant information after light interacts with matter, making their rapid and accurate sensing of great scientific significance and application value. For highly transparent materials such as living cells, variations in the transmitted light field are mainly manifested in the phase dimension, and therefore the development of highly sensitive quantitative phase imaging techniques is crucial for fast and accurate biological tissue analysis and pathological diagnosis. When light interacts with a medium, its polarization state often changes due to the dielectric properties, microstructure, and morphological characteristics of the material, and measuring the polarization state of transmitted or reflected light not only enables the perception of material properties but also enhances optical sensing capabilities in special environments such as nighttime, fog, or haze. Consequently, improving the ability to simultaneously sense multiple physical dimensions of the light field, including amplitude, phase, and polarization, has long been one of the core research directions in optical imaging. While the amplitude of a light field can be directly obtained using photoelectric detectors due to its quadratic relationship with intensity, phase and polarization information are inherently difficult to measure directly. Existing approaches for quantitative phase sensing mainly include interferometric methods, such as digital holographic interferometry, and non-interferometric phase retrieval techniques based on Hartmann wavefront sensing and the transport of intensity equation (TIE). Polarization measurement is typically realized with waveplates, polarizers, and other bulk optical components through time-division, amplitude-division, or polarization-channel-division acquisition of multiple sets of orthogonally polarized images. However, such systems are usually constrained by complex optical layouts and the size of conventional components, which hinders efficient integration and miniaturization, and in particular, time-division imaging severely limits sensing speed and temporal resolution. The emergence of optical metasurfaces provides a promising solution to these challenges, as their ability to manipulate multiple wavefront parameters of light at subwavelength scales enables substantial reduction of imaging system volume while maintaining high imaging quality. Although significant progress has been achieved in metasurface-based imaging and extensive efforts have been devoted to realizing multidimensional light-field imaging using metasurfaces, it remains challenging to develop a compact, easily integrable multidimensional imaging system that is robust and adaptable to complex imaging scenarios.

To address these issues, this work proposes a 4f optical architecture for multidimensional imaging based on a single-layer metasurface and single-shot acquisition. By employing a space- and polarization-multiplexed multi-channel metasurface, three-dimensional spatial decoupling of the orthogonal circular polarization components of the incident light field is achieved, enabling four-channel imaging with jointly controllable defocus and off-axis distances. Four intensity images are simultaneously recorded at the image plane in a single exposure, and through accurate cropping, calibration, and matching of these images, the complex amplitudes of the orthogonal circular polarization components are precisely reconstructed using the transport of intensity equation. Without the need for interferometric reference beams, the proposed architecture enables the simultaneous acquisition of intensity, phase, and polarization ellipticity information of the light field in a single exposure, significantly improving imaging efficiency and system robustness. The non-interferometric multidimensional imaging architecture is further integrated into a commercial microscope platform, demonstrating its compatibility with existing optical systems.

Quantitative phase imaging of a microlens array is performed to verify phase reconstruction accuracy, and polarization imaging of murine tendon tissue is conducted to demonstrate its capability for revealing birefringent and anisotropic structures in biological samples. Experimental results confirm that the proposed system enables label-free, fast, and multidimensional imaging of biological tissues and cells, effectively overcoming limitations of conventional imaging approaches for highly transparent samples and biological specimens, such as limited imaging dimensionality, high system complexity, and restricted applicability to in vivo measurements. Moreover, the proposed architecture introduces a system design concept that balances miniaturization, functional integration, and imaging performance, providing a practical pathway toward compact multidimensional optical imaging systems.

In principle, the proposed framework can achieve full Stokes parameter sensing based on TIE-derived phase and intensity information of orthogonal polarization components. However, in practical experiments, noise, model mismatch, and uncertainties in boundary conditions make it difficult for existing algorithms to accurately reconstruct the phase of complex tissues, especially in regions with large intensity gradients, where distortions and blurring are prone to occur, thereby hindering the reliable determination of the Stokes parameters s₁ and s₂. Future work will therefore focus on improving TIE-based phase retrieval algorithms to enhance reconstruction accuracy and robustness, as well as optimizing the matching and registration of intensity images acquired from different off-axis directions to mitigate off-axis effects, with the aim of further extending the applicability of the proposed metasurface-based multidimensional imaging architecture to complex biological, medical, and environmental sensing scenarios.

Lithium-ion batteries are critical components in modern energy systems, and their safety and reliability heavily depend on the quality of cell manufacturing. Detecting surface defects on battery cells—such as scratches, pits, wrinkles, and corrosion—is essential yet challenging due to the small size, high diversity, and complex morphology of these defects. Traditional inspection methods, including manual visual checks or conventional 2D imaging, struggle with low accuracy, high false/missed detection rates, and sensitivity to lighting variations in real-world production environments. Moreover, data acquisition under industrial conditions is often hindered by reflective surfaces (e.g., aluminum-plastic film) and inconsistent illumination. To address these issues, this study aims to develop a robust, high-precision defect detection model tailored for structured light imaging, which provides stable, high-contrast 3D-like surface information while mitigating environmental interference.

We propose YOLO-SLB, an enhanced YOLOv8 architecture specifically designed for micro-defect detection on battery cells using structured light images. The model introduces three key innovations. First, the backbone integrates CMUNeXt Blocks featuring large-kernel depthwise convolutions in an inverted bottleneck structure to strengthen spatial feature interaction and preserve fine details critical for detecting minute anomalies. Second, the neck employs a novel C2f_RFAConv module that replaces standard convolutions with receptive field attention convolution (RFAConv), which dynamically generates convolutional weights to overcome the limitations of static weight sharing in large kernels, thereby improving fusion between shallow spatial and deep semantic features. Third, the detection head is redesigned as ASFF_F—a four-branch adaptive spatial feature fusion module that not only applies learnable scale-wise weights but also adds a dedicated micro-target detection path directly utilizing high-resolution shallow features. This reconstruction enhances multi-scale perception and reduces information loss during feature pyramid operations. The model is evaluated on SLBD, a new dataset of 477 original high-resolution (1520 pixel×2000 pixel) battery cell images captured via line-scan cameras under structured light, augmented to 3147 samples covering eight common defect types.

The experimental results show that the proposed model achieves an average detection accuracy of 97.98%, with precision and recall rates of 97.05% and 92.67%, respectively, which are 4.29%, 1.97%, and 6.46% higher than those of the benchmark model YOLOv8. Extensive experiments on the NEU-DET public dataset also verified its good generalization performance, with an average detection accuracy improvement of 0.65% compared to the benchmark model, indicating that this method has high engineering application value.

YOLO-SLB effectively addresses the core challenges of micro-defect detection in lithium battery cells by synergizing structured light imaging with architectural innovations in feature extraction, fusion, and detection. The integration of CMUNeXt, C2f_RFAConv, and ASFF_F significantly boosts sensitivity to small anomalies without compromising robustness. The model's high accuracy, generalization capability, and real-time performance underscore its engineering value for automated quality control in battery manufacturing. Future work will explore semi-supervised learning and online model updating to enable continuous improvement using unlabeled production data.