GRDNet: A Lightweight Railway Foreign Object Intrusion Detection Algorithm

Railway intrusion detection using convolutional neural networks has significant results, but it requires a large amount of data and arithmetic power to train and deploy neural network models, and it is still a challenge to improve detection accuracy under limited computational resources. Based on this problem, this paper proposes a lightweight railway foreign object intrusion detection algorithm: GRDNet, which firstly adopts a lightweight backbone network GhostNetV2, and makes use of the decoupled fully connected attention mechanism to aggregate local and long-distance information, which significantly reduces the number of parameters and computation while maintaining the detection performance. Next, a deep feature extractor, RSPPFCSPC, is proposed to reduce the computational effort by decreasing the number of channels in the feature map while maintaining the diversity and effectiveness of the features. Finally, a knowledge distillation algorithm is used to improve the accuracy of foreign object localization and classification and reduce the leakage rate of the model. Experiments show that the proposed method is superior to the state-of-the-art work of each comparison.