Small Object Detection Algorithm for Railway Scene

Small object detection plays an important role in computer vision. However, in the existing methods, the detections of small objects are prone to error detection and missing detection, so the results of the methods are not ideal. When the train is running, it is necessary to detect the obstacle object at a distance. But the distant obstacle objects in the image are small and occupy fewer pixels, leading to the lack of sufficient information to achieve accurate and effective detection. Therefore, we propose a YOLOv5-based small object detection algorithm for railway scene, which adds a detection scale with quarter down-sampling. At this scale, the feature map size is larger and more favorable for small object detection. In data preprocessing, we use copy-paste small objects to increase the number of small objects, while the original images are retained during the pasting process to expand the dataset. In the experiment, our method is compared with the YOLOv5s algorithm in the railway scene dataset, and the mAP is improved by 8.4%, which verifies the effectiveness of our method.