DFSDNet: A dual-branch multi-scale feature fusion network for surface defect detection of copper strips and plates

Abstract

Surface defect detection is a research hotspot in the field of computer vision. Due to the complex characteristics of metal surfaces and the multitude of industrial defects, it remains a challenging task. In order to meet the requirement of accurate identification of surface defects on copper strips and plates in industrial quality control, we propose a computer vision-based dual-branch features fusion neural network named as DFSDNet. We gather defect samples to construct the KUST-DET dataset of surface defects on copper strips and plates to support the training and evaluation of detection models. Experiments on the KUST-DET dataset demonstrate that DFSDNet-s achieved a mean Average Precision (mAP) of 88.53%, while maintaining low computational complexity and low parameters, and the model achieves a good balance between detection precision and computational efficiency. In addition, the mAP on the NEU-DET dataset is 75.67%, showcasing its good defect detection performance. Experiments have shown that DFSDNet is an effective surface defect detection model with great potential in other metal industry applications.