Strip steel surface defect detecting method combined with a multi-layer attention mechanism network

In the production of strip steel, defect detection is a crucial step. However, current inspection techniques frequently suffer from issues like low detection accuracy and subpar real-time performance. We provide a deep learning-based strip steel surface defect detection technique to address the aforementioned issues. The algorithm is also implemented in three specific ways: as the backbone, the neck, and the detection head. Backbone employs an enhanced cross stage partial in conjunction with ResNet to effectively mine defect information and minimize the issue of adjoining feature maps' neighboring feature maps losing information. Neck is a better structure, we propose and use the dilated weighted across stages-feature pyramid network in the network to adjust the receptive field and attention weight preference of the output feature maps at different scales and to improve the utilization of defect features by the algorithm to enhance the detection of abnormal size defects. We use four detection heads in the detection head so that the network can learn the features of defects of various sizes. Finally, we use the decoupled head to separate the classification work from the regression work before combining the prediction. Two datasets of surface flaws in strip steel are used in our experiments (GC10-DET and NEU-DET). In addition, it has been shown that our proposed algorithm's mAP in GC10-DET and NEU-DET reaches 79.93% and 72.76%, respectively, resulting in a better detection impact.