Image Classification Method Based On Abandoned Stacked Restricted Boltzmann Machine

Image classification refers to the process of automatically grouping images into a set of predefined categories according to certain classification rules.It can be applied to many fields,such as face recognition and medical image processing.The traditional image classification method generally uses the artificially set feature pattern to extract features from the image,and then uses the classifier to classify the extracted features.However,for images with complex content,the artificially extracted features are difficult to meet the actual needs,which reduces the accuracy of the classification.Compared with the traditional feature-based image classification method,deep learning can realize automatic extraction of features without human intervention,but its extracted features have no corresponding mathematical or physical meaning.It is necessary to repeatedly adjust network parameters according to the classification results to extract more suitable.Characteristics.The deep belief network is a branch of deep learning,which effectively solves the problem of feature validity.However,there are too many parameters in the deep confidence network,which makes the optimization algorithm difficult to optimize and easy to fall into local optimum.Based on the research of deep learning image classification method,this paper makes an in-depth study on the application of deep confidence network model to solve image classification problem,and designs a new image classification method based on deep belief network and its network parameter optimization algorithm.In response to the above questions,the specific work of this paper is as follows:(1)An image classification network structure composed of a disposable stack-limited Boltzmann machine and a classification network is proposed.In this paper,a stack-constrained Boltzmann machine is used to form a deep confidence network.The purpose of this network is to extract suitable features from the image,and the network can verify the correctness of the extracted features.When the feature extraction is completed,the stacked constrained Boltzmann machine stops running.Features are sent to the classifier to complete the classification task.The network can avoid adjusting the parameters of the feature extraction network and the classification network at the same time,which greatly reduces the difficulty of adjusting parameters.(2)A global-local improved evolution-gradient descent algorithm is proposed,which combines the advantages of evolutionary algorithm and gradient descent method.The improved evolution-gradient descent algorithm is used as a training algorithm for classification networks to optimize the parameters of the classification network.In the improved evolution-gradient descent algorithm,the gradient descent algorithm is first used to optimize the classification network.When the gradient descent algorithm falls into stagnation,the gradient descent algorithm stops running.The improved evolutionary algorithm continues to optimize the classification network based on the gradient descent algorithm optimization.Softmax classifier.The method uses two different optimization methods to effectively reduce the probability that the optimization algorithm falls into local optimum.(3)This paper verifies the network and optimization algorithm by using mnist handwritten digital image dataset and wine dataset.The experimental results show that compared with other image classification methods,the proposed classification method has better classification accuracy and over-fitting ability.