An Improved Faster R-CNN for Devices Detection in Railway 4C System

As main devices in railway 4C system, accurate detection of insulators is essential to guarantee stable operation of whole railway network. There have been kinds of methods based on traditional image processing methods proposed in last several years. However, few of these methods meet requirement of the task due to complexity of the real environment. Deep learning has been developed rapidly in recent years and has been practically applied to computer vision domain. Detection algorithm based on these thoughts has replaced conventional methods that utilize hand-engineered features, becoming the most prevalent option in the community. Inspired by effectivity of two-stage methods, we propose a novel detection framework for detecting insulators in railway 4C system via improving the state-of-art two-stage framework Faster R-CNN. The main contribution of this framework is to apply FPN to Faster R-CNN to generate feature pyramids for both region proposal and detection. In order to obtain better performance, RoIAlign is utilized instead of RoIPool for avoiding misalignment. Experimental results reveal that this improved Faster R-CNN has remarkable performance on testing set in terms of accuracy. This framework is more robust for detecting objects at different scales, compared with original Faster R-CNN framework.