Metal sensor base defects detection using deep learning based YOLO network.

Currently, the existing object detection methods have many limitations in detecting defects on the base surface of metal sensors, such as a high rate of false detection and missed detection. Therefore, we proposed an improved algorithm based on You Look Only Once (YOLO) v5s aiming to solve the problem. Firstly, the C3 module was poor at detecting small defects. To enrich the gradient flow information and improve the detection accuracy of defects, the C2f module was used to replace part of the C3 module in the neck of the YOLO v5s. Then, an improved attention mechanism named Dilated Global Attention Mechanism was proposed to make the network focus more on the important information features. The dilated convolution was integrated into the spatial attention mechanism to enhance the receptive field of the model, reduce the model size and improve the detection performance of small defects. Finally, we proposed a novel localization loss function named Intersection over Union (IoU) with Normalized Wasserstein Distance, which not only alleviated the issue of Complete IoU loss based metrics being sensitive to the location deviations of small defects but also adjusted to diverse datasets. Results from ablation experiments demonstrated that the improved YOLO v5s algorithm enhanced the detection of the mean Average Precision by 5.3% and the Precision rate (P) by 7% compared with the original algorithm. [ABSTRACT FROM AUTHOR]