| Popis: |
Defect detection is extensively utilized within the metal industry, particularly for identifying surface imperfections on steel strips. However, the current methods still face challenges in detecting small and elongated defects on steel strips. Such defects occupy a relatively small pixel percentage within the entire image. The repeated downsampling in convolutional networks, coupled with the dynamic changes in the receptive field, can result in the potential loss of these minute defects. To mitigate the problem, our paper proposes EC-YOLO, a real-time defect detection network for steel strips of the above peculiar defects. Firstly, the 1D convolution in the efficient channel attention bottleneck (EB) module enhances the feature extraction ability of the backbone for small and elongated defects, while also facilitating the attentional mechanism for modeling channel features. Secondly, Context Transformation Networks integrate cross-stage localized blocks, referred to as CC modules, to enhance the understanding of feature semantic contextual information. Thirdly, a self-constructed dataset containing both small and elongated defects is used for understanding where such defects are more relevant in feature fusion and extraction. On the public datasets GC10-DET and NEU-DET, the improved model achieves mean Average Precision (mAP) scores of 71% and 83%, respectively, surpassing the performance of other mainstream models. The mAP of the enhanced model on the SLD-DET dataset reaches 87.5%, demonstrating its superiority in detecting both small and elongated defects. |
