Enhanced multiscale attentional feature fusion model for defect detection on steel surfaces

Abstract

Surface defect detection has been an important part of controlling the quality of industrial products. Detecting small defects poses a persistent challenge, primarily due to the scarcity of available information. To solve this problem, this paper proposes an improved model Bi-Level Efficient Global YOLO (BEG-YOLO) based on YOLOv8x. The architecture consists of Cross Stage Partial DenseNet Global Feature Pyramid Network (CSP-GFPN), and Bi-Level Efficient Attention (BEA). CSP-GFPN is a feature pyramid network structure that employs enhanced feature fusion across scales. It extends the neck network of BEG-YOLO while enabling information sharing across different spatial scales and potentially non-adjacent semantic layers. BEA is a fusion of attention. It serves to eliminate most of the irrelevant feature key–value pair inputs at the broader feature map level, thus focusing more comprehensively on the few relevant regions that remain. Three public datasets, NEU-DET, GC10-DET, and X-SDD, are used in this experiment. According to the experimental results, better results are achieved in terms of $mA{P}_{50}$, $mA{P}_{50-95}$, precision, and recall compared to other good YOLO series models. Code is available at https://github.com/xyk1300332643/BEG-YOLO.