A multiple feature-maps interaction pyramid network for defect detection of steel surface

Diverse categories, variety of shapes and uncertainty of contrast lead to challenges in accurately detecting the fine details of defects in steel surface images. Deep learning methods have provided accurate, real-time detection algorithms in the field of defect detection of steel surface in recent year. Most deep learning-based networks fuse information from each intermediate layer simply and directly, while the intrinsic relationship of feature maps with different resolutions is lacking. Therefore, a novel approach to exploit the attention mechanism, multiple feature-maps interaction pyramid network (MFIPNet), is proposed. MFIPNet is designed to consider both structural regularization and structural information in an integrated fashion by using the attention mechanism as a selector for multiple feature-maps. In MFIPNet, multiple feature-maps are integrated together with different contributions, which assists the network to better adapt to the complexity of the defects. The proposed method is evaluated on the public datasets both qualitatively and quantitatively. The results demonstrate that our method outperforms state-of-the-art approaches by more than 3.92% in mPA and 7.58% in mIoU.