Accurate sclera segmentation for biometrics remains challenging under unconstrained imaging conditions, such as varying gaze, occlusions, and inconsistent illumination. To address this, we propose GCGA-Net, a U-Net-based architecture integrating a novel global context-guided attention (GCGA) mechanism. GCGA leverages the deepest encoder features to guide the decoding process via attention-modulated skip connections. These connections are refined by modified Selective Kernel (SK) convolutions, enabling adaptive receptive-field adjustments based on learned weights. Coupled with progressive upsampling and deep supervision, this design ensures high-resolution localization and robust convergence. Extensive 5-fold cross-validation on the SBVPI, MASD.v1, MSD, and combined MASD+MSD datasets yields F_1 scores of 97.60
Reconfigurable Intelligent Surfaces that can simultaneously transmit and reflect, known as STAR-RIS, are showing great promise for boosting wireless network performance through smart manipulation of electromagnetic waves. In this study, we focus on maximizing the throughput using optimal packet size within STAR-RIS-enabled wireless communication systems that incorporate energy harvesting (EH) modules. These modules convert surrounding mechanical vibrations into usable electrical energy. We develop optimization problems designed to enhance system throughput. The numerical and theoretical results highlight the effectiveness of our proposed algorithms, demonstrating substantial improvements over existing benchmarks. This underscores the potential of vibration-based energy harvesting to support sustainable, high-throughput intelligent wireless networks.
Animal Activity Recognition (AAR) using wearable sensors is essential for precision livestock farming. While supervised deep learning advances AAR, its effectiveness is limited by the scarcity of annotated data. Self-Supervised Learning (SSL) offers a promising paradigm by extracting representations from massive unlabeled datasets. However, many prevailing self-supervised learning approaches for time series, especially contrastive methods, rely heavily on generic time-domain augmentations. Such augmentations often fail to preserve the intrinsic frequency-domain structure of physical motion, leading to representations that are sensitive to high-frequency sensor noise. To address this, we propose Spectral-Guided Contrastive Learning (SpecCL), a novel physics-informed SSL framework. First, we introduce sensor-realistic augmentations (e.g., spatial rotation and time-warping) to construct physically meaningful learning pairs simulating real-world perturbations. SpecCL is built upon a spectral-guided asymmetric encoding architecture that regularizes the latent space by incorporating a non-parametric frequency stream as a deterministic reference of spectral distributions. By minimizing the Kullback–Leibler divergence between the negative-sample similarity distributions of this fixed spectral reference and a trainable time-domain encoder, SpecCL explicitly distills physical priors into the learned temporal representations. Extensive experiments show that SpecCL achieves 90.95
As one of the common materials for parts processing and manufacturing, strip steel has wide applications in the field of machine. Due to the complexity of the processing environment, defects such as scratches, corrosion, and cracks often occur on the surface of strip steel, which are characterized by different scales, diverse types and complex backgrounds, and are highly challenging in the field of deep learning based defect detection. Compared with other network architectures, YOLOv7-tiny has the advantages of lightweight, taking into account detection accuracy and speed, which meets the requirements of real-time detection in manufacturing. Aiming at the shortcomings of YOLOv7-tiny network, such as poor feature fusion ability and low communication between layers, this paper proposes an improved YOLOv7-tiny network GD-YOLO to enhance global information fusion and multi-layer information exchange. The neck network of YOLOv7-tiny was redesigned by means of information Gather-and-Distribute Mechanism to enhance the feature information fusion capability. The multi-layer information fusion module is constructed by pooling and DySample up-sampling, and the information exchange between different layers is strengthened. The comparison experiments on data sets of NEU-DET and GC10-DET show that the performance of the proposed network model on the two data sets is improved, and the average accuracy rate (mAP@0.5) is increased by 5.4
Aerial oriented object detection confronts severe challenges arising from high aspect ratios, drastic scale variations, and complex background interference. Existing mainstream methods typically employ fixed square receptive fields for feature extraction, which constrains the model’s capability for dynamic focusing. Furthermore, complex background noise results in blurred object boundaries, thereby impeding precise localization. To address these issues, this paper proposes an Adaptive Geometry-aware Network (AG-Net) tailored for oriented object detection. First, a Multi-Scale Fusion and Dynamic Perception (MFDP) module is designed within the backbone network. Leveraging a gating mechanism, this module dynamically aggregates horizontal and vertical spatial priors with non-directional semantic information and synergizes with Multi-scale Convolution (MConv) to enhance feature representation for multi-scale and high-aspect-ratio objects. Second, the detection head is decoupled into a three-branch architecture, where a Geometry-Aware Position Refinement (GAPR) module is introduced into the regression branch. By integrating spatial and frequency domain operations, the GAPR module captures long-range spatial dependencies and edge details to improve localization accuracy amidst complex backgrounds. Finally, a Dimensionality Reduction Intersection over Union (DRIoU) loss function is designed to quantify the geometric alignment of rotated bounding boxes via multi-axis projection. Extensive experiments on the CODrone and DOTA-v1.0 datasets demonstrate that the proposed AG-Net outperforms other state-of-the-art methods.The code is available at https://github.com/zzz497/AG-Net .
