EFD-YOLOv4: A steel surface defect detection network with encoder-decoder residual block and feature alignment module

Abstract

Surface defect detection is often required in industrial production process. However, due to the large variation in defect scale, low contrast between the background and defect target, and noise interference, the task of defect detection in industrial environments is still challenging. This paper proposes an enhanced-feature-detection YOLOv4 model for steel surface defect detection, named EFD-YOLOv4. Firstly, an augmented path consisting of a convolutional encoder-decoder module is developed for the residual block to enhance the learned representations. A multi-scale module based on hierarchical residual-like connections is applied to enlarge the receptive fields. Then, a feature alignment module with an attention mechanism is designed for feature misalignment in feature fusion. Finally, three decoupled heads are adopted to output classification and regression results independently. The experimental results show that this method achieves 79.88 mAP on the NEU-DET dataset and 54.65 mAP on the GC10-DET dataset. The proposed method has better performance on steel surface defect detection tasks.