| Deep convolution neural network has achieved great success in image processing,semantic understanding and other applications.However,due to the large number of parameters and large amount of computation in deep convolutional neural network,it is not suitable for embedding directly into mobile applications.Therefore,it is necessary to study the compression method of the parametric model and optimize the computation in the convolutional neural network.The weight parameters of the deep convolutional neural network is analyzed,and based on the traditional weight attenuation pruning method,we proposed a compression method of neural network model for pruning recurrently.By using the Rectified Linear Unit(ReLU)activation function and L1 norm regularization method,the neural network could be sparser.Then,we use the three tuple vector to represent the sparse matrix.Finally,we use sparse matrix multiplication instead of convolution operation to reduce the amount of computation in neural network.We design and implement a target object capture application of iPhone operation system.This application can capture objects offline,and broadcast location information by voice.The main contents and results of this paper are as follows:(1)In order to solve the problem of excessive weight parameters in the model of deep convolutional neural network,the compression method of pruning recurrently is proposed.Pruning recurrently means pruning few connections every time,then retraining.And repeat this pruning progress.Experiments on MNIST dataset and CIFAR-10 dataset show that the method has a higher accuracy than the traditional pruning method when the number of pruned connections is more than 95%of the total connections.(2)For the large quantity calculation in convolutional neural network,we carry out some optimization based on sparse matrix operation.First,we make the neural network become sparser by using the ReLU activation function and L1 norm regularization method.And then we use the three tuple vector to represent the sparse matrix.Finally we use sparse matrix multiplication instead of convolution operation to reduce the amount of calculation in neural network.(3)The target object capture application is designed and implemented based on the compressional neural network model and the convolution optimization method.The compressed neural network model is embedded in the mobile application,and the convolution optimization method is combined to make the neural network model run smoothly at the mobile system.The overall architecture of the application is designed,and the target object name acquisition module,the pixel information acquisition module,the object recognition module and the capture result processing module are designed and implemented in detail.Then we evaluate this application on function and performance.In this paper,we mainly study the compression method of deep convolutional neural network.Then combining the compression method and the convolution optimization method,we design and implement the application of target object capture,and evaluate the application. |
