Steel Surface Defect Detection Algorithm Based on YOLOv8

• DOI: 10.3390/electronics13050988
• Publication Date: 2024-03-05T16:26:59Z
• Authors: Xuan Song; Shuzhen Cao; Jingwei Zhang; Zhenguo Hou
• Keywords: 0203 mechanical engineering; 02 engineering and technology

To improve the accuracy of steel surface defect detection, an improved model multi-directional optimization based on YOLOv8 algorithm was proposed in this study. First, we innovate CSP Bottleneck with two convolutions (C2F) module by introducing deformable convolution (DCN) technology to enhance learning and expression ability complex texture irregular shape features. Secondly, advanced Bidirectional Feature Pyramid Network (BiFPN) structure is adopted realize weight distribution input features different scales feature fusion stage, allowing for more effective integration multi-level information. Next, BiFormer attention mechanism embedded backbone network, adaptively allocate target features, such as flexibly efficiently skipping non-critical areas, focusing identifying potentially defective parts. Finally, adjusted loss function from Complete-Intersection over Union (CIoU) Wise-IoUv3 (WIoUv3) used its dynamic non-monotony property effectively solve problem overfitting low quality bounding box. The experimental results show that mean Average Precision (mAP) task detection reaches 84.8%, which depicts a significant improvement 6.9% compared original YOLO8 model. can quickly accurately locate classify all kinds defects practical applications meet needs industrial production.