Research On Image Classification Algorithm Based On Deep Learning

With the continuous upgrading and popularization of mobile image acquisition devices such as mobile phones,image data has gradually become the most important source of daily information acquisition.However,with the increasing scale of image data,redundant information is more and more full of people's side,which makes it more and more difficult for people to search for useful information,so the image recognition and automatic classification has more and more important significance.With the application of deep learning in image classification in recent years,more and more excellent image classification structures based on deep convolution neural network have achieved remarkable results.However,the image classification method based on deep learning is more dependent on large data sets,resulting in a large amount of parameter calculation in the process of image classification,which leads to the problems of long classification process and low efficiency of classifier.The two most important parts of image classification process are feature extraction part and classifier design part: on the one hand,feature extraction is the basis of all image classification tasks.It is very important to study efficient feature extraction algorithm in the image field,and good feature extraction can greatly improve the effect of image classification.On the other hand,the design of classifier directly determines the classification effect of image classification structure.In conclusion,from the perspective of feature extraction and classifier design,this paper aims to solve the problems of long classification process and low efficiency of classifier.The research work is as follows:The dimensionality reduction principle of Principal Component Analysis(PCA)for image feature data,and the calculation of covariance matrix between feature data and its eigenvalues and corresponding eigenvectors in the process of dimensionality reduction have been studied.In this paper,the image entropy is introduced to optimize the dimensionality reduction process of PCA.Before the dimensionality reduction of PCA,the image entropy is first used to conduct the preliminary screening of image features,so as to reduce the image feature scale and reduce the redundant information of the image.Then,the features after the preliminary screening are further dimensionalized.Experiments show that the optimized PCA dimension-reduction algorithm can not only greatly reduce the time of image feature extraction and selection,but also improve the quality of feature extraction to a certain extent.The feasibility and significance of the Extreme Learning Machine(ELM)as a Convolutional Neural Network Network(CNN)classifier is studied in this paper.To solve the problem that deep learning does not fully utilize image features in image feature extraction,a new image classification method based on feature fusion is proposed.The hybrid model utilizes CNN to effectively extract global image features and SIFT local features for fusion,and reduces the dimension of fused features.Meanwhile,it utilizes the fast and efficient feature of ELM classifier to enable the two methods to work together.Experiments show that the algorithm not only ensures the classification accuracy,but also greatly reduces the training time of deep learning algorithm.