YOLOv5-ACCOF Steel Surface Defect Detection Algorithm

Abstract

Steel surface defect detection is critical for industrial production and quality control. Traditional methods, however, face challenges such as complex scenes and the detection of small defects. To address these issues, we propose a surface defect detection method based on the YOLOv5-ACCOF architecture. Our approach incorporates the Convolutional Block Attention Module (CBAM) into the Backbone core feature extraction module C3, thereby enhancing the focus on key information within the Backbone layer. We replace the conventional SPPF layer with Atrous Spatial Pyramid Pooling (ASPP), improving the model’s capability to perceive multi-scale information. The Content-Aware ReAssembly of Features (CARAFE) module substitutes the upsampling module in the Neck layer, enabling content-aware feature reassembly, thereby improving detail retention in the feature maps and enhancing the model’s upsampling efficacy. Additionally, we integrate Omni-Directional Dynamic Convolution (ODConv) into the C3 module of the Neck, which facilitates richer feature representation through multi-directional convolution operations, thereby augmenting the model’s defect detection capability in complex backgrounds. The loss function is modified to Focaler-IoU, which improves the detector’s performance across various detection tasks by focusing on different regression samples. Experimental results demonstrate that our proposed YOLOv5-ACCOF model significantly outperforms the original model, exhibiting robust generalization capabilities, thus verifying its effectiveness and feasibility in practical applications.