The Metal Defect Detection Model Based on Non-Local Fusion Attention and Residual Networks

In response to the challenges in extracting features, low detection accuracy, and large parameter sizes in current metal surface defect detection, a metal defect detection model based on non-local fusion attention and residual networks was designed. First, Ghost convolution was used to replace traditional convolution modules to reduce the network parameters. Second, to address the issue of defect information loss during feature extraction in deep networks, a fusion of residual networks and defect downsampling was proposed, resulting in the Residual Downsampling Network (RDNet). Finally, to enhance the feature extraction capability of the model, the Non-local Fusion Attention Network (NFANet) was introduced, which effectively utilizes global information to improve defect detection accuracy. The model achieved average precision (AP) of 99.8% and 87.9% on the publicly available metal defect datasets KSDD and KSDD2, respectively. The rate of missed and false detections significantly decreased, and the extracted defect features became more refined, achieving high detection performance while keeping the model's parameter size relatively small.