Surface Defect Detection of Steel Strip with Double Pyramid Network

Defect detection on the surface of steel strip is essential for quality assurance strip. Precise localization and classification, two significant tasks defect detection, still need to be completed due diversity scales. In this paper, a residual atrous spatial pyramid pooling (RASPP) module first designed enrich multi-scale information feature maps increase receptive field maps. Secondly, double network (DPN) that combines RASPP proposed fuse features further so similar semantic are shared among each layer. Finally, DPN-Detector, an automatic defects network, proposed, which embeds DPN into Faster R-CNN replaces original head with head. Experiments carried out dataset (NEU-DET), results show mAP DPN-Detector as high 80.93%, 3.52% higher than baseline R-CNN. The classification accuracy 74.64%, speed reaches 18.62 FPS. method performs better robustness, regression capability other methods.