A lightweight steel surface defect detection network based on YOLOv9
In the steel production process, surface defect detection is crucial for ensuring product quality. To address the issues of high computational cost and low detection accuracy in current steel defect detection models, we propose a YOLOv9-based steel defect detection algorithm, CCSS-YOLO. First, we in...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
AIP Publishing LLC
2025-05-01
|
| Series: | AIP Advances |
| Online Access: | http://dx.doi.org/10.1063/5.0273824 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | In the steel production process, surface defect detection is crucial for ensuring product quality. To address the issues of high computational cost and low detection accuracy in current steel defect detection models, we propose a YOLOv9-based steel defect detection algorithm, CCSS-YOLO. First, we integrate the SENetV2 attention mechanism, which has a multi-branch structure and compression and excitation operations, into the YOLOv9 feature module RepNCSPELAN4 to form a new RepNCSPELAN4_SENetV2 feature extraction module, enhancing the model’s ability to extract complex features. Next, we replace the regular convolution blocks in the model network with spatial-to-depth convolutions, further reducing the model’s computational complexity while retaining global feature information. Finally, we replace the Fusion module in the CNN-based cross-scale feature fusion (CCFM) module, with the new Fusion-RepNCSPELAN4 module, creating a new feature fusion network, CCFM-YOLO, which replaces the neck network of YOLOv9. This approach improves the model’s feature extraction capability while reducing its parameter count. Experimental results show that, on the NEU-DET dataset, CCSS-YOLO improves average detection accuracy by 1.9% compared to the original YOLOv9 model, reduces parameters by 25.4%, and decreases FLOPs by 15.6%. In addition, it demonstrates good generalization performance on the new ASVD washer dataset. |
|---|---|
| ISSN: | 2158-3226 |