YOLOXD: A New Network for Metal Surface Defect Detection

The detection of metal surface defects is an essential part of the industrial production process. Most of the current detection methods are poorly detected due to the diversity of metal surface defect morphologies and the complex reality of the environment. To address this problem, this paper proposes a novel metal surface defect detection network YOLOXD based on YOLOX, which is designed with Dilated Pyramid Pooling to increase the receptive field and introduces an attention mechanism to enhance feature extraction compared to YOLOX. And the Mish activation function is used to assist the model training. Our method achieves 78.45%
$\mathbf{mAP}$
in the proposed dataset.