Global attention module and cascade fusion network for steel surface defect detection

Steel surface defect detection plays a pivotal role in contemporary society, ensuring quality and safety in construction and manufacturing, reducing production costs, improving efficiency, and driving technological innovation. However, this task encounters challenges, including addressing unstructured features, multi-scale issues, and a scarcity of available data. To overcome these challenges, this paper proposes a global attention module and cascade fusion network for steel surface defect detection, called GC-Net. In this network, the global attention module is proposed to enhance the capability of the model to handle unstructured defects. Subsequently, a cascade fusion network is designed for multi-scale feature fusion, thereby improving detection accuracy for defects of varying scales. Following this, soft non-maximum suppression is applied in the post-processing stage to eliminate redundant detection boxes, further enhancing the detection performance of the network. Finally, a series of data augmentation techniques, including oversampling and small object augmentation, are employed in the experimental sessions to mitigate the issue of data scarcity. The experimental results on two datasets for steel surface defect detection demonstrate that the proposed method outperforms state-of-the-art methods in terms of mAP50 metric (NEU-DET: 0.771, GC10-DET: 0.635). The code is released at https://github.com/Ghlerrix/GC-Net. © 2024 Elsevier Ltd