Image Classification Method Based On Deep Learning And Accelerated Training Technique

With the rapid development of the Internet and the popularity of terminals such as smartphones,image data on the Internet has shown an exponential growth trend.These images cover all aspects of human life and contain a large amount of useful information.Image classification is an important means to use this information,which is helpful for solving many problems in realistic scene.It has broad application prospects in image retrieval,scene recognition and human-computer interaction.The convolutional neural network is widely used in the field of image classification because of its excellent feature extraction capabilities.As the actual task scenarios become more and more complex,the scale of the network model is also increasing in order to meet the demand,leading to trained a network model consumes a lot of time.Therefore,how to design an efficient convolutional neural network and how to accelerate the training of neural networks has become an important research topic in the field of deep learning.In this thesis,the image classification is used as the background,and the typical convolutional neural network DenseNet is selected as the research object.The method of image classification using DenseNet and the distributed parallel training algorithm are studied.DenseNet reduces the parameter quantity of the model through feature multiplexing,and can achieve high accuracy,but the scalability of the model is not high,and the frequent feature map concatenating operation also makes the model take up too much video memory in the training process.Synchronous data parallelism is the most commonly used parallel training method.In each iterative process,communication nodes need to communicate with the parameter server to exchange gradient values.When the network model parameter size is large or the network bandwidth is limited,the communication time is likely to become a bottleneck,which greatly affects the training speed of the model.This thesis first proposes an improvement strategy for the shortcomings of the DenseNet model.By changing the way of feature reuse in the model,the model parameters are reduced,and three different convolutional neural networks are designed based on the improved model.They are trained on MNIST and CIFAR-10 data sets respectively,achieving high accuracy and realizing image classification.The experimental results show that the improved strategy proposed in this thesis can significantly reduce the parameters of the model and improve the efficiency of the parameters under the premise of achieving the same accuracy.Secondly,aiming at the communication bottleneck problem in synchronous data parallelism,an optimization algorithm based on gradient quantization is proposed.The algorithm overlaps a part of the calculation time and the communication time by updating the parameters layer by layer.Before the calculation nodes upload the gradient information to the parameter server,the gradient value is quantized and encoded to reduce the amount of communication data.In order to reduce the influence of the gradient loss caused by quantization on the convergence speed of the model,the quantization error is added when each node updates the parameters,and the parameters of each node are averaged after several iterations.The experimental results show that compared with the traditional synchronous data parallelism algorithm and SSP algorithm,the optimization algorithm can effectively improve the training speed of the model and alleviate the communication bottleneck problem of distributed training.