Steel Surface Defect Detection Method Based on Improved YOLOv9 Network

Abstract

This paper introduces a new steel defect detection model CK-NET, which uses YOLOv9c as the baseline model and adopts YOLOv9's model architecture for improvement. The proposed model addresses issues of shallow information loss and insufficient feature extraction and fusion caused by network deepening. A new feature extraction module is designed to control model parameters and enhance feature extraction. Minor improvements to Convolutional Block of Attention Module (CBAM) and the introduction of deformable convolutions in the backbone network further enhance feature extraction. A new feature fusion module, combined with a self-attention mechanism (SA), elegantly fuses features from different levels to assist downstream detection tasks. The Programmable Gradient Information (PGI) auxiliary branch is also improved to better fuse features and guide model learning with gradient information. All improved modules have been integrated into CK-NET. Experiments on the NEU-DET dataset demonstrate that CK-Net achieves a 13.2% higher mAP value than YOLOv9c, reaching a 92.1% mAP value while maintaining similar model parameters, validating the model's effectiveness.