Multi-scale Rotated Bounding Box-Based Deep Learning Method for Electric Railway Detection

摘要

This paper introduces a rotation-based framework for Electric Railway detection in patrol inspection scene images, concretely for droppers under different shooting angles. The rotated detection network which is used for generating inclined proposals with dropper orientation information, is particularly designed for long anchor boxes settings and RoI filtering criteria. And then the subsequent defect network applied for judging the state of droppers bases on the cropped local image pixels that exclude excess background information. On the foundation of this frame, defect net woks more intently on dropper pixels in the region proposals. We conduct experiments using this rotated frame under more complex shooting conditions, e.g. complicated background, droppers close to each other. The results demonstrate the rotated detection is more robust under various situations and the defect discrimination responses more focus on the droppers only.