A three-stage framework for multi-type pantograph anomaly detection under complex environments

A novel three-stage framework is proposed in this paper for detecting pantograph anomalies. This framework is capable of detecting anomalies in multiple types of pantographs and is resilient to complex backgrounds and illumination variations, exhibiting strong robustness. In the first stage, the improved Yolov8 network is utilized to localize the pantograph region, addressing the issue of complex backgrounds during pantograph detection. In the second stage, the Short-Term Dense Concatenate (STDC) network is employed for precise segmentation of the pantograph region. Furthermore, corresponding improvements are made to the network to handle edge blurring caused by illumination variations. In the third stage, binary images of different types of pantographs are transformed into vectors that contain pantograph features. Additionally, Relief-F and random forest algorithms are employed for feature selection and anomaly classification. Ultimately, the proposed framework achieves an average accuracy of 97.04% for various anomaly types in a testing set consisting of images of multiple types of pantographs.