Research On Image Classification Methods Based On Deep Learning Models

In recent years,with the rapid development of Internet technology,the amount of data on the Internet has also increased dramatically,and the technology of data collection has also greatly improved.Moreover,more and more people use images to represent and transmit information.In order to people get information quickly and accurately,the Image Classification has been widely studied as an important field of artificial intelligence.In addition,the key to Image Recognition problem lies in how to extract more abstract feature information from the image.Therefore,the feature extraction is very important for the results.Deep learning is a multilayer neural network learning algorithm that has emerged in recent years,and it is widely used in different fields with its excellent feature learning ability,such as image classification,speech recognition,and natural language processing.Firstly,on the basis of Image Classification and Deep Learning in this paper,two classic Deep Learning models have been studied in depth: Deep Belief Network(DBN)and Convolutional Neural Network(CNN).Secondly,Aiming at the problem of slow convergence and poor classification accuracy due to fixed learning parameters in the training process of traditional deep belief network.In this paper,there is an AML-LBP-DBN algorithm has been proposed based on adaptive momentum and learning rate.Firstly,the adaptive update criterion was introduced into the training of Restricted Boltzmann Machines;secondly,adaptive adjustment of momentum and learning rate is achieved by reconstructing the error increments and updating the weights before and after iteration;finally,ORL and Yale face database were used to verify the validity of the algorithm.Compared with the LBP-DBN algorithm,Experimental results show that AML-LBP-DBN algorithm can obtain faster convergence speed and higher classification accuracy.Finally,Aiming at the problem of the poor classification accuracy based on Convolutional Neural Network(CNN),caused by the random initialization of weights during the training process and the noise interference in process of feature extracting,this paper presents an integrated optimization method of simulated annealing(SA)and Gaussian convolution.Firstly,the algorithm extracts the central feature as priori information,find the optimal solution as initial weights of full-connection layer by simulated annealing,and then accelerate the weight updating and convergence rate;secondly,smoothing the image by Gaussian convolution to reduce noise disturbing;finally,the integrated network is applied to the MNIST and CIFAR-10 databases,and it can be found that accuracy rate of the integrated network is improved through the contrastive analysis of different algorithms.