The rapid development of industrialization has led to a significant increase in the demand for steel, making the detection of surface defects in steel a critical challenge in industrial quality control. These defects exhibit diverse morphological characteristics and complex patterns, which pose substantial challenges to traditional detection models, particularly regarding multi-scale feature extraction and information retention across network depths. To address these limitations, we propose the Dynamic Sparse Attention Transformer (DSAT), a novel architecture that integrates two key innovations: (1) a Dynamic Sparse Attention (DSA) mechanism, which adaptively focuses on defect-salient regions while minimizing computational overhead; (2) an enhanced SPPF-GhostConv module, which combines Spatial Pyramid Pooling Fast with Ghost Convolution to achieve efficient hierarchical feature fusion. Extensive experimental evaluations on the NEU-DET and GC10-DE datasets demonstrate the superior performance of our approach. On the NEU-DET dataset, DSAT achieved an accuracy of 92.55%, with a mean Average Precision (mAP) of 83.14% at an IoU of 0.5 and 47.37% at an IoU of 0.5:0.95, significantly outperforming existing methods. Similar improvements were observed on the GC10-DE dataset, where the model attained an accuracy of 79.61% and a mAP of 67.27% at an IoU of 0.5, and a mAP of 34.09% at an IoU of 0.5:0.95. Comprehensive ablation studies validated the respective contributions of the DSA mechanism and the SPPF-GhostConv module, while visualization experiments demonstrated the model's enhanced capability in detecting fine-grained defects. Overall, the proposed DSAT model exhibits significant advantages and potential in the field of steel surface defect detection.
The detection of defects on steel surfaces constitutes a vital area of research in computer vision, characterized by its complexity and variety, which pose significant difficulties for accurate identification. In this context, we introduce a deep learning framework that combines multi-channel random coding with modules for multi-scale feature fusion to tackle the challenges of low recognition accuracy and insufficient classification power prevalent in conventional models. Our model capitalizes on the self-attention mechanism associated with the Transformer architecture, alongside the strong feature extraction capabilities of Convolutional Neural Networks (CNNs), to facilitate a combined improvement in performance. To start, we enhance the model’s feature extraction functionality by incorporating ResNet18 along with global self-attention. Next, we bring forth a novel improvement to the backbone network by adding a multi-channel shuffled encoding module, which effectively encodes various features through the interactions of different convolutional groups, thus minimizing the number of parameters. Additionally, we introduce a multi-feature fusion module UPC-SimAM (upsample concatenated Simple Parameter-Free Attention Module), which is free from parameter redundancy to bolster the model’s capacity to merge multi-scale features. Our experiments on the NEU-DET and GC10-DE datasets demonstrate that our model outperforms existing state-of-the-art techniques regarding detection efficiency. Specifically, the model registers a classification accuracy of 91.72%, an mAP@0.5 of 83.03%, and an mAP@0.5:0.95 of 45.55% on the NEU-DET dataset. On the GC10-DE dataset, it achieves a classification precision of 76.73%, an mAP@0.5 of 65.03%, and an mAP@0.5:0.95 of 32.46%. Through detailed ablation studies and visualization experiments, we affirm the considerable potential and benefits of the proposed SH-DETR model in the field of detecting defects on steel surfaces.
Rice diseases pose a severe threat to global food security, while traditional detection methods suffer from low efficiency and dependence on manual expertise. To address the challenges of insufficient feature extraction and poor multi-scale disease adaptability in existing deep learning approaches under complex field environments, this study proposes ADAM-DETR, a rice disease detection algorithm based on improved RT-DETR. We constructed the RiDDET-5 dataset containing 9,303 images covering five major disease categories. The algorithm innovatively designs three core modules: the AdaptiveVision Network (AVN) backbone for enhanced feature extraction, the Dual-Domain Enhanced Transformer (DDET) module for spatiotemporal-frequency domain collaboration, and the Adaptive Multi-scale Feature Model (AMFM) for improved feature fusion. Experimental results demonstrate that ADAM-DETR achieves 94.76% mAP@50 on the RiDDET-5 dataset, representing a 3.25% improvement over the baseline, and 83.32% mAP@50 on the public Kamatis dataset with a 2.19% enhancement, validating its cross-domain generalization capability. The algorithm requires only 42.8G FLOPs with 14.3M parameters, achieving an optimal balance between accuracy and efficiency, providing an effective technical solution for disease monitoring in smart agriculture.
Remote sensing images are extensively utilized for Earth observation; however, they often encounter challenges such as extremely small objects, cluttered backgrounds, and constraints related to edge computing. Current studies frequently face difficulties in achieving an optimal balance between recognition accuracy and efficiency to satisfy the requirements of various embedded devices. We propose DFE-DETR, a lightweight dynamic feature-enhanced detector that refines RT-DETR for remote-sensing scenes. The model relies on three core modules: the Sparse Attention Enhancement Module (SAEM) that prunes irrelevant tokens via Top-k scoring to cut compute, the Multi-scale Convolutional Attention Enhancement Module (MSCAEM) that employs multi-branch depth-wise strip convolutions to sharpen slender targets, and the Deformable Large Kernel Convolution Module (DLKCM) that uses adaptive sampling to locate irregular objects. Experiments demonstrate that this framework achieves a significant performance breakthrough on the VisDrone2019 dataset with only 23.6 M parameters, reaching 47.34% mAP@0.5 and 65.8 FPS. On the SIMD dataset, it attains 82.79% mAP@0.5 and 74.6 FPS. Meanwhile, on the RSOD and NWPU VHR-10 datasets, it achieves mAP@0.5 scores of 96.58% and 93.28%, respectively. Ablation experiments validate the synergistic effects of the modules, and visualization heatmaps indicate that the proposed model provides the most comprehensive target perception, accurately delineating target contours. In comparative experiments, the model outperformed mainstream YOLO series and specialized UAV detection models, offering a new high-precision real-time detection solution for embedded scenarios such as disaster response and traffic monitoring.
11 October 2025 in Materials Science and Engineering: C