Recurrent attention convolutional neural network optimise track foreign body detection

Abstract

Abstract Timely and accurate identification of track foreign body intrusion is of great significance for maintaining the safety of railway running. Object detection technology based on deep learning can be used to judge the foreign body in the track running area. However, the general target detection technology cannot recognise the foreign object information in complex environment, which has a certain hysteresis. In this study, adaptive adjustment is made to the algorithm according to the limited area in the railway special scene. A new Recurrent Attention Convolutional Neural Network and adaptive particle swarm fusion network is proposed. Produced in the process of image classification and regional sampling information loss, by introducing attention study under the multi‐scale recursive fusion, each scale is a network of target recognition and attention network, starting from the full image, continuously put forward attention area, and as input to realise detection to the next dimension positioning; With the help of adaptive particle swarm optimisation algorithm, the feature information at different levels is fused, and the high‐attention region is generated by continuous recursive clipping from rough to fine, The mutual enhancement of target area detection and fine‐grained features is realised. Experiments conducted by the author in Railway data set showed that its target detection accuracy reached 84.2%. Compared with other existing models, the relative accuracy gain was 10.6% and 8.1%, respectively, and the maximum gain was 6.2% compared with the self‐ablation experiment.