A Dense Attention Railway Foreign Object Detection Algorithm Based on Mask R-CNN

In railway safety monitoring, foreign object detection is a key task, especially in infrared and low illumination conditions. In order to improve the detection accuracy of railway foreign bodies, an optimized railway foreign bodies detection algorithm named O-Mask R-CNN is proposed in this study. Firstly, by integrating the densely connected feature pyramid network (FPN) and the convolutional attention mechanism (CBAM),the recognition ability of low-contrast objects and the detection accuracy of small-size foreign objects are significantly improved. In addition, O-Mask R-CNN uses an improved Zone Suggestion Network (RPN) and ROIAlign layer to ensure feature alignment, thereby optimizing the precision of the bounding box and the quality of the instance segmentation. Finally, a method of adjusting the size of anchor frame based on cluster analysis is introduced to adapt to the characteristics of different scale railway foreign bodies. After only 46 iterations, the developed railway foreign item identification algorithm achieved a constant total loss value and has strong iterative performance. Mean square error and mean absolute error are the smallest, 1.35 and 1.21 respectively, and the detection error performance is also very good. Finally,the real average detection accuracy of the algorithm on infrared and common railway foreign body photos is tested, which is up to 98.26 and 98.85, respectively. Therefore, the detection method created in this paper not only has high performance, but also provides technical support for foreign object recognition in practical application environment.