Multi-scale and dynamic snake convolution-based YOLOv9 for steel surface defect detection: Multi-scale and dynamic snake...: J. Chen et al.

Steel surface defect detection plays a critical role in ensuring the quality of steel production. However, challenges such as low contrast in defect visibility and the irregularity and variability of defect patterns often affect the performance of existing detection methods. To address these challenges, we propose a novel approach based on multi-scale and dynamic snake convolution-based YOLOv9 (MDSC-YOLOv9) for steel surface defect detection. By leveraging multi-scale and dynamic snake convolutions, MDSC-YOLOv9 has strong feature representative ability and sensitivity to narrow and elongated defect patterns. We introduce an adaptive multi-scale Retinex with color restoration for image preprocessing, which enhances the content of image information, generating more diverse and robust training data. Furthermore, the locally enhanced positional encoding (LePE) attention module and the upsampling Dysample module are integrated into the multi-scale and dynamic snake convolution feature fusion network of MDSC-YOLOv9 to further enhance the feature information of small targets and reduce the information loss during upsampling. Experimental results on both the NEU-DET and GC10-DET datasets demonstrate that our approach outperforms existing methods in terms of accuracy and has strong generalization ability across diverse datasets. [ABSTRACT FROM AUTHOR]