Study On Light-weight Algorithm Of Deep Neural Network For Railway Foreign Object Intrusion Detection

Foreign object intrusion detection based on video intelligent analysis is an important technical means in the field of railway perimeter safety monitoring and prevention,which is of great significance to ensure the safety of railway operation.In recent years,with the rapid development of deep learning,video analysis algorithms based on convolutional neural network and other deep neural networks have greatly improved the detection performance.However,due to the problems of high complexity,long training time,large amount of calculation and storage,and high demand for hardware,deep neural networks cannot be widely used in the railway perimeter security system with a large number of surveillance cameras.Therefore,this paper studies a series of problems such as convolutional neural network pruning,compression training,automatic optimization of neural architecture,and explores the light-weight methods of deep neural networks,so that the railway video intelligent detection algorithm can better meet the practical application requirements of less hardware resources,real-time and high reliability.Firstly,aiming at the problem that the existing neural network redundancy identification criteria can not accurately judge the redundant convolutional kernels,this paper proposes a redundancy identification criterion based on the feature map norm,and evaluates the importance of the convolutional kernels by calculating the feature map norm.In order to further improve the identification degree of redundant convolutional kernels in each layer,the identification criterion based on the layer-wise feature map norm is proposed,and the recursive pruning is performed on the convolutional layer,so as to reduce the accuracy loss while gradually compressing the model.Experimental results on several public datasets including CIFAR,SVHN and Image Net show that compared with the existing redundancy identification criteria based on convolutional kernel norm,the proposed algorithm has the ability of more accurate identification of redundant convolutional kernels,and can obtain higher recognition accuracy under the same compression rate,which lays important foundations for subsequent structured sparsity and fast compression of networks.Aiming at the problem that existing pruning compression algorithms generally require multiple recursive pruning-retraining,the paper proposes the fast compression algorithm for convolutional neural networks based on structured feature sparsity training,which adds the structured feature norm regularization to the conventional training loss function The regularization term promotes the feature channels sparsity and layers sparsity,and realizes the single-step pruning of the network under the condition of stable accuracy.Different from other existing pruning algorithms,structured feature sparsity training avoids multiple rounds of recursive pruning-retraining steps,improves the efficiency of model compression training,and is conducive to the rapid completion of independent compression and optimization of deep neural network in each monitoring camera along the railway.Experimental results on multiple public classification datasets show that the proposed algorithm can enhance the saliency of the remaining feature channels,reduce the time-consuming network compression training,and outperform the existing state-of-the-art pruning methods.Aiming at the problems that the existing pruning compression algorithm is limited by the width of the pre-training large network and the pruning hyperparameters require manual intervention,the dynamic channel scaling with feature sparsity learning and similarity measurement is proposed.The algorithm optimizes the network through the three targets of increasing network consumption cost,feature sparsity constraint and similarity measurement,and realizes the end-to-end search of the neural architecture under specified resource conditions.The experimental results on several public datasets show that compared with the existing light-weight network and other compression algorithms,the neural network searched by the proposed algorithm can obtain higher recognition accuracy under the condition of less storage space,which lays the foundation for the subsequent light-weight detection algorithm.Finally,aiming at the requirement of fast and reliable multi-scale target recognition in railway scene,the network search space including dynamic receptive field operation and measurable dense residual connection is proposed to realize multi-scale feature extraction and measurable cross-layer feature information transmission.The efficient and light-weight network is obtained by using the search algorithm proposed in this paper,which is used as the backbone of the target detector to realize the light-weight target detection algorithm.In the application of rapid recognition of foreign object intrusion in high-speed railway,using mixed scene training of multi cameras and single camera scene transfer training,the proposed light-weight detection algorithm can achieve higher detection performance while occupying lower hardware resources.Compared with the existing technologies,the series of network light-weight algorithms proposed in this paper can greatly compress and accelerate the neural network,so that the deep neural network can be applied to the online identification and monitoring of high-speed railway perimeter intrusion,and reduce the application cost of intelligent security system.At present,it has been applied in many high-speed railways.