Edge feature enhancement: Generating adversarial edge perturbations for preterm infant movement recognition.

ABSTRACT Next‐generation intelligent perception systems require photodetectors that are functionally reconfigurable for in‐sensor processing. However, integrating broadband response, polarization sensitivity, and electrically reconfigurable photoresponse into a single compact device remains challenging. Here, we demonstrate a solution using a gate‐programmable photodetector based on a 2H‐MoTe 2 /ReSe 2 heterojunction. The device exhibits self‐powered broadband detection from 300 to 900 nm, with a responsivity of 0.945 A/W, a specific detectivity of 1.41 × 10 12 Jones, and a polarization ratio of 16.45 at 635 nm. Critically, gate voltage asymmetrically modulates the dual built‐in fields via Fermi‐level tuning, enabling continuous and programmatic reconfiguration of the magnitude and polarity of photocurrent. The responsivity is tuned from 0.142 to −1.254 A/W, and this polarity reversal allows the polarization ratio range to span both positive and negative domains. Leveraging this unique programmability, we realize two advanced applications: multimodal encryption communication, where gate voltage and polarization angle serve as dual dynamic encryption keys, and in‐sensor image convolution for direct edge extraction at the sensor level. This work establishes a robust platform toward multifunctional and reconfigurable optoelectronics, advancing the development of all‐in‐one intelligent perception.

ABSTRACT Amidst the rapid evolution of the information society, there is a pressing demand for photodetectors which integrate high performance with imaging and processing capabilities. However, implementing traditional image processing relying on predefined convolution kernels inherently suffers from constrained adaptability and is difficult in individual optimization. Here, we introduce a self‐rolled‐up Te/Graphene photothermoelectric detector (STGP), pioneering a physics‐response‐driven self‐construct convolution kernel strategy rooted in its wide‐angle detection performance for advanced image feature extraction. The STGP leverages a tubular morphology that efficiently enhances light absorption via light‐trapping effect while the built‐in electric field is established at the Te/Graphene interface due to their work function difference, which promotes the directional transport of photo‐thermally excited carriers, substantially suppressing interfacial recombination losses. The STGP realizes a high voltage responsivity of 259 V W −1 at 940 nm and an ultrafast response time. Functionally, the unique tubular geometry enables exceptional wide‐angle detection, which is used as the basis for self‐construct convolution kernels and achieve effective image sharpening and edge extraction. This work not only provides a novel built‐in field enhancement strategy for photothermoelectric detectors but also establishes a new paradigm for in‐sensor computing by integrating the material's physical response with computational functions.

Metasurfaces are confined to static functionalities and lack reconfigurability-a key characteristic urgently needed for their practical applications in dynamic environments. To address the critical challenges of traditional metasurfaces, including fixed functions, polarization dependence, bulky imaging systems, difficulties in integrating edge detection with bright-field imaging, and the requirement for additional digital post-processing, we propose to leverage the dynamic reconfigurability enabled by phase change materials, combining it with polarization insensitivity and omnidirectional dynamic switching between high-resolution edge extraction and clear bright-field imaging. In this paper, we propose a dual-polarization Laplacian differentiator operating in the terahertz band based on a nonlocal perforated metasurface, with dynamic function switching achieved by regulating the phase transition of vanadium dioxide (VO2). When VO2 is in the insulating state, the device can directly perform two-dimensional second-order image edge detection. When VO2 transitions to the metallic state, it switches to bright-field imaging mode. The Optical Transfer Function (OTF) required for Laplacian operations is achieved by exciting the Quasi-Bound States in the Continuum (Q-BIC) mode under p- and s-polarized terahertz wave illumination, which endows the device with an angular dispersive response matching the Laplacian operator's requirements. This differentiator offers dual-polarization-compatible edge detection, and its efficient, high-performance function switching-coupled with the benefits of dual-polarization imaging-provides robust technical support for terahertz-band applications including machine vision, biomedical detection, and image processing.

Although optoelectronic memristors with nonvolatile bipolar photoconductivity enable in-sensor vision-centric neuromorphic hardware, achieving wavelength-defined polarity inversion across a broad spectrum remains a challenging task. Herein, a stable optoelectronic memristor composed of nonstoichiometric lead oxide (PbOx) coated black phosphorus (BP) nanosheets is demonstrated. The optoelectronic processes in the PbOx-BP heterostructure result in programmable polar photoresponses across the 365nm - 1,550nm wavelength range. Visible light causes positive photoconductance via photoelectrochemical Ag+ reduction and conductive filament reconstruction. Conversely, ultraviolet light drives the reverse photogenerated electron transfer to chemically oxidize the Ag CFs, while infrared light induces their localized melting via the photothermal effect. This bipolar optoelectronic tunability enables all-optical Boolean logic operations, allowing for the realization of 14 binary functions through optical reconfiguration. Furthermore, multispectral computing tasks, including edge extraction and spectral noise suppression, are performed, yielding a classification accuracy of up to 98.6% for 16 crop species using an all-optical convolutional neural network. The ultra-thin oxide coating presents an effective surface modification approach to improve two-dimensional devices, while the optoelectronic bipolarity establishes a framework for all-optical modulation in neuromorphic machine vision.

Automatic crack segmentation is crucial for ensuring the safe and stable operation of civil concrete buildings. However, due to the irregularity of cracks, low image quality, and complex background environment, automatic crack segmentation on concrete building surfaces still faces significant challenges. To address these issues, an automatic segmentation network (LKT-Net) based on a large kernel pooling Transformer is proposed, aiming to improve the comprehensiveness and accuracy of crack feature extraction while maintaining a lightweight design. First, the large kernel pooling Transformer (LKT) is proposed as the fundamental building block of LKT-Net, which combines large kernel convolution with pooling layers and attention mechanisms to effectively enhance global perception and capture local details at a lower computational cost. To extract edge information accurately, the Feedforward network is improved by integrating the Laplacian operator with multi-scale convolutions, thereby enhancing multiscale edge detection capabilities. Finally, to mitigate information loss during downsampling, we propose a feature enhancement module (FEM) to replace traditional skip-connections, thereby enhancing cross-level feature interactions. The experimental results showed that on three public datasets (DeepCrack537, CrackLS315, and CrackTree260), compared with eight advanced networks, LKT-Net achieved mean Intersection over Union (mIoU) scores of 86.23%, 70.82%, and 83.67%, respectively, demonstrating excellent segmentation performance. The codes are available at: https://github.com/wjxcsust2024/LKT-Net .

Abstract Pointer meters are widely used in industrial environments due to their inherent stability and reliability. To enhance the feature extraction capability and edge contour segmentation accuracy of automatic reading methods, this article proposes Pointer Meter Scale Aggregation Network (PMSANet), a novel architecture for accurate pointer meters reading through hierarchical multi-scale feature aggregation. First, PMSANet uses SA-Basicblock as the backbone network to classify pointer meters in natural environments, where the introduction of the attention mechanism solves the problem of fixed feature weight allocation in traditional methods. Second, the MS-UNet, which is based on multi-scale feature aggregation, is utilized as the segmentation network to accurately segment the pointer, main scale lines, and center point of the dial. The multi-scale feature aggregation can fully capture semantic information and edge details of features at different levels, improving the network’s feature extraction capability. Finally, automatic meter reading is achieved through the angle method. Compared with advanced segmentation networks on our dataset multi-category pointer meter dataset (MCPMD) and public dataset (Pointer-10K), PMSANet demonstrates the best performance. Compared with the U-Netv2 method, it improved Dice by 4.42%, Intersection over Union by 7.48%, and mean absolute error by as little as 0.003. The reading results show that the overall maximum error rate does not exceed 0.981% and the average error rate does not exceed 0.674% on the MCPMD. The method proposed in this article can meet the practical application requirements for automatic pointer meter reading accuracy.

Infrared small target detection remains challenging in applications such as long-range surveillance and early warning due to the fact that infrared images rely on thermal radiation, which results in limited texture cues and a low signal-to-noise ratio for the targets. Although recent deep networks have improved representation capability, they often exhibit two persistent limitations. Fine target details are gradually weakened through successive downsampling, and edge-related priors are not sufficiently exploited to stabilize target responses under background interference. To alleviate these issues, an Edge-Prior Guided Dual-Branch Enhancement Network (EGDENet) is proposed, a dual-branch framework that injects edge priors into feature learning for infrared small target detection. An auxiliary edge-aware branch is introduced to complement the main encoder–decoder stream. Specifically, a Multi-directional Sobel Edge Extraction (MSEE) module is designed to adaptively reweight multi-directional edge responses, thereby strengthening boundary-sensitive representations. Furthermore, a Difference-Aware Gated Fusion (DAGF) module leverages Gated Spatial Convolution to capture subtle variations in the features and employs depthwise separable convolution along with adaptive enhancement to effectively integrate the extracted edge information. In addition, an Edge Pixel Integration (EPI) Loss is present to couple edge sensitivity with pixel-wise supervision. This loss improves the edge sensitivity of infrared small targets. The proposed EGDENet is evaluated on three benchmark datasets: NUAA-SIRST, IRSTD-1K, and SIRST-Aug. The experimental results show that our method outperforms or matches the performance of state-of-the-art methods.

To improve sketch recognition accuracy, this study proposes an enhanced sketch recognition model based on an improved CycleGAN network and a dual attention mechanism. The proposed model first incorporates multi-directional convolution and brightness equalization modules into the CycleGAN network to extract edge and contour features. A dual attention mechanism is then implemented using channel attention and spatial attention modules, effectively addressing issues of sparse strokes and uneven spatial distribution in sketches while enhancing the representation of critical features. Finally, a hybrid architecture combining global average pooling and convolution layers serves as the classifier to produce sketch recognition results. Simulation results demonstrate that this model achieves 97.08% accuracy, 98.12% precision, 98.23% recall, and 97.45% F1 score for sketch recognition on the TU-Berlin dataset, and 98.65% accuracy, 98.12% precision, 98.76% recall, and 97.95% F1 score on the QuickDraw dataset. Compared with state-of-the-art sketch recognition models, this model exhibits superior performance in accuracy. These results indicate that the model can enhance sketch recognition precision and provide technical support for converting sketches into high-quality animated images.

ABSTRACT Quantitative differential phase contrast (qDPC) microscopy enables high‐resolution, label‐free imaging of weakly absorbing samples by combining asymmetrical illumination with phase transfer function (PTF) deconvolution. However, conventional methods are limited by the ill‐posed nature of deconvolution and the band‐limited characteristics of the PTF, leading to poor robustness against noise and background fluctuations, particularly in thick or complex samples. To overcome these challenges, we propose a pupil‐driven differential phase contrast (PD‐DPC) framework that integrates system PTFs into both the data fidelity and regularization terms of the reconstruction model. The proposed model incorporates an edge‐sparsity‐promoting regularization to preserve structural detail and suppress noise, along with a Retinex‐inspired fidelity formulation to mitigate background fluctuations. The resulting non‐convex optimization problem is solved via an efficient Split Bregman algorithm with iterative reweighted soft‐thresholding. Simulations and experiments demonstrate that PD‐DPC outperforms L2‐DPC, Iso‐DPC, TV‐DPC, and Retinex‐DPC in terms of background suppression, phase fidelity, and edge preservation. The framework is compatible with diverse DPC modalities and enables automatic cell contour segmentation as well as high‐resolution imaging of absorbing tissues beyond the weak‐object approximation. By combining physics‐informed priors with a data‐adaptive reconstruction strategy, PD‐DPC offers a robust, broadly applicable solution that substantially enhances the accuracy and applicability of qDPC for biomedical imaging. The MATLAB code is available on GitHub .

This paper proposes a novel salient object detection method for operational hole localization in metallurgical furnaces, addressing challenging industrial conditions including extreme illumination variations and strong electromagnetic interference to enable two-level measurement in aluminum electrolysis cells and impact position recognition of the front-of-furnace operation robot. It employs a multi-feature fusion framework combining foreground and background saliency maps with center prior maps. Foreground saliency maps are generated through spatial compactness and local contrast computations, enhancing discriminative features while suppressing shared foreground–background characteristics. Background saliency maps are constructed via sparse reconstruction to exploit redundant features. Then method integrates edge extraction and density clustering to generate center prior maps that emphasize foreground target centroids and mitigate background noise. Comprehensive evaluations on both a specialized operational hole dataset and six public datasets demonstrate superior performance compared to other methods. On the specialized dataset, it achieves a precision of 0.8954, a maximum F-measure of 0.8994, and an S-measure of 0.8662. While maintaining operational robustness, the method offers a practical solution for furnace monitoring and robotic operation guidance in metallurgical processes.

An early and accurate diagnosis of coronary artery disease (CAD) is essential for effective treatment. Although X-ray coronary angiography (XCA) is the clinical gold standard for CAD diagnosis, blurred vessel boundaries, low contrast, and minute stenotic regions can make pixel-level segmentation difficult. We propose Stenosis-YOLO, a YOLOv8-based segmentation framework that addresses these challenges. Its key contributions are as follows: (1) an edge enhancement stem (EES) that combines Laplacian edge extraction with spatial feature pooling to strengthen fine vascular and boundary representations; (2) a small-target-aware neck that uses space-to-depth convolution (SPD-Conv) on the P2 layer and a feature fusion module (FFM) to improve multiscale integration; and (3) a semisupervised pseudo-label self-training strategy that uses unlabeled data to improve performance. When evaluated on the ARCADE benchmark, Stenosis-YOLO significantly outperformed the state-of-the-art techniques for coronary stenosis segmentation and instance segmentation, achieving improvements of 6.4%, 4.9%, 5.7%, and 4.2% in terms of precision, recall, F1-score, and mean average precision (mAP), respectively, over the baseline YOLOv8 model. Stenosis-YOLO also performed exceptionally well in stenosis detection, achieving 0.976, 0.972, 0.974, and 0.983 for precision, recall, F1-score, and mAP, respectively. This represents enhancements of 1.6% and 0.7% in the F1-score and mAP, respectively, compared to the leading coronary stenosis detection model. These results demonstrate the effectiveness of combining edge enhancement, small-target modeling, and semisupervised learning for accurate coronary stenosis segmentation.

Abstract Underwater imaging is often degraded by wavelength-dependent absorption and light scattering, leading to color distortion, multiscale blur, and spatially non-uniform attenuation. Such degradations not only reduce visual quality but also undermine the stability of feature extraction that many underwater visual measurement pipelines rely on. Existing methods often face challenges in maintaining robustness and preserving measurement-relevant structural cues under such complex degradation. We propose a dynamic graph and multi-head latent attention collaborative network (DGMLA-Net). Specifically, the dynamic graph feature enhancement module captures spatial dependencies to suppress noise and correct non-uniform degradations, while the low-rank efficient attention module enables global context aggregation and cross-scale feature fusion with reduced memory overhead. Experiments on multiple underwater benchmarks demonstrate that DGMLA-Net achieves competitive performance on standard image-quality metrics. More importantly, measurement-oriented evaluations show that the enhanced images improve the reliability of downstream feature processing, yielding more stable keypoint detection and more continuous edge extraction. These results indicate that DGMLA-Net provides feature-preserving underwater imagery as measurement-oriented preprocessing for subsequent vision-based measurement tasks.

The integration of artificial intelligence into image-guided intraoperative interventions holds considerable promise for deriving 3D geometric information from 2D imaging. 2D/3D registration establishes the spatial relationship between preoperative computed tomography (CT) and intraoperative X-rays. However, existing methods are often limited by the image domain gap and imprecise feature extraction, causing coarse registration to provide inadequate initial poses and subsequent fine registration to fall into local optima, thereby reducingaccuracy. We aim to develop a robust single-view lumbar spine 2D/3D registration framework that balances high clinical accuracy with intraoperative efficiency requirements by aligning preoperative CT with intraoperativeX-rays. We propose utilizing vertebral body edges in X-rays as novel semantic features to guide 2D/3D registration. For robust edge extraction, we develop ESegMamba, an efficient U-shaped Mamba network incorporating Group multi-axis Hadamard Product Attention (GHPA) and Group Aggregation Concatenation (GAC) modules. Experiments for semantic edge extraction were performed on a dataset of 710 images (comprising X-rays and Digitally Reconstructed Radiographs) derived from 10 patients. The dataset was partitioned using a 4:1 patient-specific split, resulting in 568 training and 142 test images. The training set was further utilized via 5-fold cross-validation for network fine-tuning. ESegMamba was benchmarked against SegMamba, SwinUNETR, and UNETR using Dice and mIoU metrics. For 2D/3D registration, experiments were conducted separately on 300 simulated samples and 90 real clinical samples, following the same patient-specific split. The proposed framework was compared with landmark-based, intensity-based, and learning-based methods using mean Target Registration Error (mTRE). Statistical significance was assessed using the Wilcoxon signed-rank test with a significance level of 0.05, applying Bonferroni correction for multiplecomparisons. ESegMamba outperforms representative networks with fewer parameters (99.18 M), achieving 90.36% Dice and 85.49% mIoU on the test set. Compared to the strong baseline SegMamba, ESegMamba demonstrated a large effect size in Dice improvement (Cohen's , ). For 2D/3D registration, the proposed method demonstrated superior performance over representative benchmarks. Specifically, compared to Xreg and PSSS, our method achieved large practical improvements in mTRE ( and , respectively; ). On real clinical data, the method achieved a mean in-plane translation error of approximately 1.5 mm and an average registration time of approximately 10s. The proposed method, empowered by ESegMamba, yields statistically significant improvements over intensity-based benchmarks ( ). The achieved sub-2mm accuracy and 10 s processing time on clinical data confirm its efficacy for intraoperative spinal navigation. The code for the proposed method is available at github.com/shenao1995/lineReg.

An efficient edge extraction method for large hollow thin-walled components based on echolocation

Adaptive frequency collaboration for remote sensing change detection.

Optical computing presents a promising avenue to meet the escalating computational demands. However, optical analog computing is susceptible to environmental perturbations, relies heavily on digital-to-analog converters and analog-to-digital converters, and requires electronic or photonic nonlinear operations. While optical digital computing mitigates some issues, its reliance on manual, task-specific configuration hinders broader applications like inference. Here, we propose the concept of an optical logic convolutional neural network (OLCNN). We demonstrate a 1-by-3 optical logic convolutional operator (OLCO) for pattern generation and validate its high-speed computing capacity at 20 Gbit/s. A 2-by-2 OLCO is then implemented to perform three types of image edge extraction. By scaling up, a 3-by-3 OLCO is constructed for an OLCNN to achieve four-class classification on the MNIST dataset with an average test accuracy of 95.1%. By synergizing optical logic devices with neural networks, this work pioneers a logic-driven paradigm for high-speed, energy-efficient optical hardware in artificial intelligence.

Underwater wireless power transfer (UWPT) technology can improve the endurance of unmanned underwater vehicles (UUVs). The stability and efficiency of UWPT depend on the success rate of UUV docking. A novel detection model, TFDF-YOLO, is proposed for dynamic position identification of UUV docking. First, a spatial–frequency decoupling (SFD) module is proposed by using Fourier-based degradation cues to guide Top-K proxy attention to boost blurred edge extraction capability. A relevance-difference fusion (RD-Fusion) strategy is improved by a global channel attention mechanism to realize multi-scale feature recognition. Furthermore, a new adaptive loss function (U-CIoU) is developed to suppress illumination bias and anchor inflation. Results on a reliable multi-source dataset demonstrate that the proposed model achieves 91.5% accuracy and 92.7% mAP@0.5. This work could enhance the success rate and reliability of UWPT. It shows potential for broader underwater applications, including deep-sea docking and multi-AUV cooperative systems.

Intrusion of foreign objects into the Electrified Railway Catenary System can lead to power failures, train service interruptions, and even casualties, making accurate detection essential for safe operation. Due to the scarcity of railway datasets, this study constructs a Railway Catenary Foreign Object Dataset to support model training and evaluation. Existing detection methods often struggle with complex railway environments, diverse object morphologies, and varying scales. To address these challenges, we propose a Railway Catenary Foreign Object Detection Network. It leverages the hierarchical architecture and window-based attention mechanism of Swin Transformer for multi-scale semantic feature extraction and global relational modeling, effectively distinguishing foreground from background. A Multi-branch Fusion Feature Pyramid Network is designed to deeply fuse low- and high-level features across scales, improving detection of objects of different sizes. Additionally, a Regional Receptive Field-Enhanced Edge Module expands the receptive field and enhances edge extraction for elongated foreign objects. Extensive experiments on the constructed dataset demonstrate the effectiveness of the proposed approach, achieving an Average Precision of 60.2%, with 53.8% for small object detection.

Real-time perception of bearing rotation angles is essential for structural health assessment of bridges. However, existing vision-based rotation angle measurement methods exhibit limited robustness to time-varying operational conditions and tracking errors, particularly in practical applications of bridge monitoring. To address this limitation, this study presents an advanced computer vision-based monitoring technology for bridge bearing rotation angles by incorporating specifically configured retroreflective targets, an efficient target tracking approach, and a rotation angle calculation algorithm. Firstly, under LED illumination, retroreflective targets appear as bright, high-contrast features in the images, facilitating precise detection and tracking. Secondly, target centroids are tracked with sub-pixel accuracy through thresholding, edge extraction, and ellipse fitting. Lastly, the bearing rotation angle is calculated by analyzing the angle between the two characteristic lines formed by the target centroids. To validate the effectiveness of the proposed method, comprehensive numerical investigations were conducted, and the results showed that the proposed method maintained high accuracy across various imaging conditions. Additionally, comparative analysis with an existing advanced method also revealed that the proposed method exhibits superior measurement performance even under target tracking uncertainties. To investigate its feasibility and validate its practical effectiveness, a field application on an 80 m + 80 m continuous beam was conducted, and minute rotation angle measurements during 23 railway train drive-by events were obtained using the proposed method, yielding a root mean square error of 0.0008° and mean absolute error of 0.0007°. The successful development and field deployment demonstrate significant potential for advancing structural health monitoring technologies, contributing to intelligent infrastructure management through automated monitoring and early warning capabilities.