An Improved Feature Pyramid Network and Metric Learning Approach for Rail Surface Defect Detection

When deep learning methods are used to detect rail surface defects, the training accuracy declines due small defects and an insufficient number of samples. This paper investigates problem defect detection by using improved feature pyramid network (FPN) metric approach. Firstly, FPN is adding deformable convolution convolutional block attention modules improve detecting different scales, it pretrained on MS COCO dataset. Secondly, a new model established extract features based transfer learned parameters. Thirdly, multimodal structure constructed, distance between each modal representative embedded vector calculated classify defects. Finally, experiments carried out miniImageNet dataset The results show that mAP (five-way five-shot) our method 73.42% 63.29% Our effectiveness proposed method, satisfactory. As there few sample classification studies this work provides approach lays foundation for further research.