Research On Image Classification Method Based On Neural Network

As one of the deep learning algorithms,convolutional neural network is an application of deep learning about image processing.The advantage of convolutional neural network is that it can directly convolve with image pixels and extract image features from image pixels.This method is closer to the processing method of the human brain visual system.The weight sharing attribute and the pooling layer of the convolutional neural network greatly reduce the parameters that the network needs to train,simplify the network model,and improve the efficiency of training.In recent years,it has become a relatively active research direction in the field of computer vision and artificial intelligence.The main work of the thesis is to study the accuracy and time cost of the image classification algorithm,and carry out in-depth research from the two aspects of the overall network design of the convolutional neural network and the optimization of the hyperparameters.A new convolutional neural network Mynet with a total of four major convolutional layer frameworks was designed.The main innovations are as follows:1)Drawing on the Inception framework,it solves the problem that the convolution kernel used to extract features in the general convolutional neural network is too single and does not have the diversity and comprehensiveness.2)The Inception architecture and the Resnet layer are combined to propose a new structure,which solves the Inception framework and increase the cost of the operation,increase the network training time,and reduces the overall efficiency of the network.3)A new data cleaning method is proposed to solve the problem of data imbalance.The convolutional neural network is implemented by the traditional deep learning framework caffe.The performance of deep learning algorithm in image classification is verified by comparing with other traditional classification algorithms.The experimental results show that the use of Mynet Convolutional Neural Network(CNN)for image classification and recognition has the advantages of high precision and fast calculation speed,and does not need to consider the design feature extraction algorithm,which has strong generalization ability.At last,the accuracy of the Mynet network is verified by using two kinds of pictures,and the stability of the network is indicated by the confusion matrix.Provides reference recommendations for image classification technology research and computer vision.