Aiming at the problems of low segmentation accuracy and low detection efficiency of split pin defect detection algorithms based on semantic segmentation, this paper proposes an improved method of split pin defect detection for catenary based on DeepLabv3+. Firstly, the MobileNetv2 network is pruned, and the MobileNetv2 network is replaced with Xception for feature extraction, which can greatly reduce the consumption of computing resources and improve the detection efficiency. Then, CBAM(Convolutional Block Attention Module) is integrated into ASPP(Atrous Spatial Pyramid Pooling) module, and CBAM is introduced to process shallow features of input Decoder network, enhance the perception of split pin edge features, and improve the accuracy of model semantic segmentation. In order to alleviate the negative impact caused by the imbalance between split pin region and background region and improve split pin segmentation accuracy, CEDice Loss is used as the Loss function in this paper, combining the advantages of Cross -Entropy Loss and Dice Loss. Finally, according to the principle of split pin defect discrimination formulated in this paper, the state recognition of split pin is carried out according to the color and shape information of image segmentation. The experimental results show that compared with the original DeepLabv3+ model, the MPA and MIOU of the improved DeepLabv3+ model are improved by 3.54% and 3.42%, respectively, and the testing time is reduced by 14.41 ms per image, and the model parameters is reduced by 88.61%. In terms of defect identification, the accuracy of the method for missing, loose and normal split pins is 100%, 98.1% and 99.5%, respectively, which can quickly and effectively identify split pin defects.