The Lightweight Steel Surface Defect Detection Algorithm Based on YOLO-ETG

To address the current challenges of complex structures, inadequate detection accuracy, and excessive para meters in steel surface defect detection models, this paper pr oposes a lightweight detection algorithm, YOLO-ETG (YOLOv8-EfficientNet, Triplet attention and ghostconv). The algor ithm utilizes EfficientNet in the feature extraction stage to r educe model parameters and introduces a low-parameter Tri plet-attention mechanism in the feature fusion phase to enha nce the capture of semantic information related to defect fea tures. Additionally, Ghostconv is employed for convolution o ptimization, further reducing parameters and computational load. In the prediction phase, the loss function is optimized t o WIOU to improve the model's convergence efficiency and speed. Experimental results show that YOLO-ETG outperfor ms YOLOv8n on the NEU-DET dataset, with a 300% reducti on in model parameters and a 69% reduction in floating-poi nt operations, achieving a mean average precision (mAP) of 78.1% and a 1% improvement in detection accuracy. This d emonstrates that YOLO-ETG effectively enhances the precisi on of steel defect detection while maintaining lightweight cha racteristics, meeting practical production needs.