Optimization And Application Of Convolutional Neural Network Algorithm For Image Recognition

The Convolutional Neural Network(CNN)is one of the representative algorithms of deep learning.CNN extracts features through convolution,and reduces the order of magnitude of network parameters through weight sharing and pooling,and finally completes tasks by classification through traditional neural networks.It has good fault tolerance,self-learning ability,self-adaptation,and can automatically extract features,so it has been widely used in image recognition,target tracking and other fields.In this thesis,the traditional convolutional neural network algorithm is optimized,and the improved algorithm is applied to image recognition tasks,and the effectiveness of the algorithm is verified.The main work of this thesis is as follows:1.The initial value of the convolution kernel and the number of feature maps in the traditional convolution neural network is difficult to determine.In order to solve this problem,an optimized convolution neural network depth learning algorithm based on unsupervised pre-training and grey relational analysis(GRA)was proposed.Firstly,the training dataset without labels was used to train a sparse autoencoder and get good initial values of convolutional filters with characteristics in accord with the input pictures.To choose a better number of convolution layer's feature maps,GRA method was introduced.The operation can remove the sample data with less relation between feature maps and the output,so that the system can automatically select the hidden layer feature maps with great influence on optimizing the network structure and improving the recognition accuracy of the system.Simulation experiments were based on global handwritten digital database MNIST.The experimental results show that compared with the traditional convolution neural network,the optimized algorithm both improved the recognition rate and recognition speed.2.As an important part of the convolutional neural network,the pooled model extracts representative values in the pooled domain by numerical calculation,so it is crucial to the accuracy of the model.The maximum pool and average pooling take the maximum and average values in the pooled domain as the results are not very representative.So,in order to further improve the accuracy of the convolutional neural network model and optimize the learning performance of the model,this paper proposes an improved pooling model based on maximum pooling and average pooling,and the global handwritten digital datasets MNIST and CIFAR-10 data.The effectiveness of the improved pooling model was verified on the two dataset.Comparing with the common pooling model,it is found that the learning performance of the convolutional neural network with improved pooling model is better.In one iteration,the error rate decreases by 4.28% on the MNIST and decreases by 2.15%on CIFAR-10 datasets.3.As the depth of the convolutional neural network increases and the structure is perfected,the performance of the model is getting better and better.However,at the same time,the storage space required for the model and the time required for prediction are greatly increased,so the computing power of hardware resources is more dependent.For the deep learning recognition task of the mobile terminal,it is very important to design a task based on the deep learning model to be able to quickly and efficiently complete the task using only the hardware resources of the mobile terminal.This thesis proposes an improved Mobilenet algorithm for the problem of low algorithm performance that the deep face recognition task encounters.The supervised signal of Mobilenet algorithm is improved to AM-Softmax from Softmax.Through many experiments,AM-Softmax is designed to be more suitable for Mobilenet algorithm with suitable additional margin and scale value.The training set and verification set are derived from the dataset MS-Celeb-1M-v1 c and dataset Asian-Celeb,and the effectiveness of the improved Mobilenet algorithm is verified on the LFW data set.Compared with the initial Mobilenet algorithm model,it is found that the performance of the improved mobilenet algorithm is better,and the accuracy rate is increased by 10% compared with softmax.By making full use of the Asian celebrity ID in the dataset Asian-Celeb to increase the number of training samples,performance is further improved by four percentage points.