A Scale-Adaptive and Background-Robust Method for Surface Defect Detection

Abstract

Despite deep learning-based methods perform remarkably well in surface defect detection recently, the unpredictable shapes and sizes of surface defects and complicated texture background still pose enormous challenges for existing methods. To address these problems, we propose a novel surface defect detection method based on YOLOv5 that combines multi-branch attention mechanism, multi-scale effective feature fusion and a lightweight neck structure, dubbed SABR-YOLO. Firstly, we embed Dual-Branch Convolution Module (DBCM) to the backbone network to enhance feature representations. Secondly, Multi-Scale Spatial Fusion (MSSF) is designed to fuse features with different scales effectively. Thirdly, we propose an Attention-Based Upsampling (ABUp) module to capture local features and focus on subtle defects. In addition, we design the Context-Aware Feature Pyramid (CAFP) by applying the ABUp and making improvements to neck structure. Finally, we evaluate our framework SABR-YOLO on the NEU-DET and HRIPCB datasets, and experimental results show that our method achieves higher accuracy compared to the state-of-the-art methods.