AFF-Net: A Strip Steel Surface Defect Detection Network via Adaptive Focusing Features

In strip steel production, detecting surface defects is crucial for ensuring product quality and optimizing production line efficiency. However, detecting defects is complicated by the variations in size, complex structures, and the wide range of defect morphologies present in strip steel. To tackle these challenges, this paper proposes a strip steel surface defect detection network via adaptive focusing features (AFF-Net). Firstly, an adaptive focusing feature block (AFF-Block) is proposed, which applies the "Diffusion-Aggregation" thought. This block repositions and adaptively assigns weights to defect features, guiding the network to focus on defect features and more effectively capture defects' spatial and morphological changes. Subsequently, a focused feature pyramid network (Foc-FPN) is proposed to enhance the network's adaptability to complex defects through multi-scale focusing fusion. This innovative structure adaptively balances the semantic gap of defect features at different scales and alleviates the abstraction feature overload. The proposed algorithm achieved a mean Average Precision (mAP@IoU=0.5) of 83.5% on the public NEU-DET dataset for strip steel surface defects, surpassing the baseline network by 8.2%. Compared to existing models, this detection method strengthens the connection between defect characteristics and more effectively detects irregularly distributed defects in complex strip steel surface images.