AMFNet: aggregated multi-level feature interaction fusion network for defect detection on steel surfaces

Abstract

Current models for detecting defects on steel surfaces struggle to fully utilize potential positional and semantic information. Usually, these models merely combine high-level and low-level features in a straightforward manner, leading to an increase in redundant information. To address this challenge, this study presents an aggregated multi-level feature interaction fusion network (AMFNet). Specifically, the AMFNet incorporates a branch interaction module (BIM) that branches and fuses features channel-wise to facilitate feature interaction. Moreover, it also includes a dilated context module (DCM) that expands the receptive field to capture contextual information across various scales effectively. In addition, we propose a spatial correlation module (SCM) that is designed to recognize spatial dependencies between adjacent feature maps and generate attention weights. Our performance evaluations on the NEU-DET and GC10-DET dataset reveal that our proposed AMFNet significantly outperforms classical object detectors in terms of mean average precision (mAP). Moreover, it also demonstrates a modest improvement over the advanced methods recently introduced by other researchers.