Fast Training Algorithm Of Classification Network For Railway Perimeter Intrusion Based On Transfer Compression

The intrusion of obstacles into the railway perimeter poses a potential threat to the operation safety of the railway system.The perimeter intrusion detection system based on video analysis needs to complete the identification and classification of foreign objects intrusion,so as to timely alarm to ensure the safety of operation.Considering the complicated railway scenes,many environmental interference factors,and the multi-camera collaboration method for monitoring,it is of great significance to study the railway video intelligent identification technology that can adapt to various scenes and has high classification accuracy with good cost performance.At present,the deep learning algorithm has been well applied in the railway scene,but there are still multiple problems such as weak classification effect,poor generalization performance,long training time,large space occupied by model memory and so on,which is difficult to meet the actual needs.Aiming at this problem,this thesis studies a deep neural network training algorithm based on network transfer compression.It uses transfer learning to efficiently optimize deep neural networks of different scenes,and performs synchronous pruning compression about the networks to solve the problem of large network parameters and large memory usage.Firstly,the abundant sample image database by collecting the existing multi-camera railway surveillance videos is established in the thesis,including three categories of empty scene,train operation,and personnel intrusion.Then,the problem of large computational complexity of the existing VGG16 convolutional network model is improved,and an optimal integrated classification network model is trained on a database of multiple camera scenes.Next,aiming at the problem that the optimal integrated classification network model has low identification accuracy and large network scale on a single camera scene,the method is proposed that network transfer learning and network compression are simultaneously conducted,which further optimize network structure and parameters on the small sample of a single camera scene.For the shortcomings of existing network compression criteria,a recursive pruning criterion based on the L1 norm or L2 norm of feature maps is proposed.By repeatedly pruning the redundant convolution kernels,the network is continuously compressed to meet the target requirement.Finally,the classification effects of different camera scenes are tested by the method proposed in this thesis.The experimental results show that the method can achieve large compression of the original VGG16 network model,and the loss of precision is small,which satisfies the rapid training requirements of a large number of deep identification networks for different scenes.