GFF-Net: A high-accuracy railway foreign body detection network based on gradient optimization and cross-layer feature fusion

Abstract The emergence of foreign bodies on railways is characterized by strong randomness and significant concealment, making high-accuracy detection a challenging task. This paper proposes a detection network based on gradient optimization and cross-layer feature fusion (GFF-Net) for the high-accuracy detection of foreign bodies on railways. Firstly, the backbone is redesigned through gradient design to optimize the feature extraction process, improving object feature representation ability while increasing feature extraction efficiency. Secondly, a cross-layer feature fusion method is designed to reduce feature loss in the traditional fusion process and enhance the model's multi-scale feature fusion capability. Then, based on the railway foreign body detection dataset (RFDD) established in this paper, GFF-Net achieves the highest detection accuracy of 89.8% mAP and a real-time inference speed of 94 FPS on the NVIDIA GeForce GTX1080Ti GPU platform, outperforming other methods like YOLOv7. Additionally, GFF-Net also demonstrates satisfactory detection performance on the public dataset MS COCO 2017.