Research On Compact Deep Learning Methods For Track Scene Classification

High-speed railway development has led the world in our country,the railway daily detection is the key of the protection of the normal operation of high-speed railway,with the continuous upgrading of track inspection car as well as the rapid development of computer vision technology,using computer vision technology instead of artificial test automation is developing unceasingly,compared with the traditional manual detection methods,automatic detection method of detection based on computer vision technology is faster and more efficient.Scene classification can promote fasteners in trajectory detection and rail track important part such as the detection performance of disease detection,with the global wave of deep learning,and deep learning become the preferred method of image classification,but it is used widely in high performance computing capacity of neural network on the equipment demand is high,the training model of reasoning is slow,it’s hard to do real time detection;Although compact neural network requires low computing power and slow reasoning speed of the trained model,its performance cannot meet the requirements.Track scene images are divided into four categories: ballastless turnout,ballastless lead line,ballastless lead line,ballastless lead line,ballastless lead line,ballastless lead line,ballastless lead line,ballastless lead line,ballastless lead line,ballastless lead line,ballastless lead line,ballastless lead line and ballastless lead line.Drawing on the design idea of the most advanced convolutional neural network in recent years,we customized the network according to the features of our orbital scene classification with few categories of tasks,small differences among classes and high consistency of pictures,and designed Dwres Net and Dw Attention Net of residual network with attentional mechanism with deep feature reuse.Feature-reusable deep separable network Dwres Net is better than most of the commonly used high-performance neural networks and compact neural networks in recognition accuracy,inference speed,model size,and computational load,and can well solve the orbit scene classification task.However,it is easy to misidentify a picture with a stern line as a sparse turnout.In order to solve the problem that the featurereusable deep separable network does not have a high recognition rate for a sparse turnout picture,an attention mechanism is introduced to increase the amount of calculation with a small amount of calculation The performance of the model is improved,and the accuracy of the orbit scene classification task is further improved.To verify the effectiveness of our method,we used the same training method to train more than a dozen models,including the commonly used high performance neural network and compact neural network.The experimental results show that,compared with several commonly used compact neural networks,The reasoning speed of Dwres Net is faster than that of all neural networks,and its performance is better than that of compact neural networks and some high-performance neural networks.