Chinese Painting Image Classification Research Based On Deep Learning

With the development of digital technology,it is significantly meaningful to realize the accurate classification and fast searching of Chinese painting for the establishment of painting images system.The traditional Chinese painting recognition system contains two main steps,which are feature extraction and classification.Feature extraction is based on personal experiences,which leads to two problems including loss of detail information and low generalization ability of math model.To discover an automatic and efficiency recognized technology will be a research hotspot in the future.Painting classification,as one type of images classification,its research difficulties come from the feature extraction.Deep learning methods could learn better feature representation through training large amounts of data compared with traditional shallow learning model,which improves accuracy of classification in many applications.Therefore,in view of the problems of Chinese painting image classification,this paper studies the Chinese painting image classification algorithm based on depth learning.In this paper,a method based on DBN(Deep Belief Network)is firstly proposed to classify the Chinese painting.On the one hand,the CDBN model takes the two-dimensional structure information ignored in DBN model into account.On the other hand,it uses the convolutional operation,which makes the CDBN model be robust to noise and then learns good features containing statistical properties.At the same time,we introduce probability model and sparse regularization into CRBM(Convolutional Restricted Boltzmann Machine),which improves the ability of reasoning of CRBM model and weakens the model over-complete,and further enhances the performance of painting classification.Then,we classify Chinese painting images based on a CNN(Convolutional Neural Network).We introduce an improved SMOTE(Synthetic Minority Over-sampling Technique)to overcome the over-fitting problem.The technique preprocesses the input raw data and then directly input the amplified data into CNN model,after the convolution and sub sampling of hidden layer,we replace traditional sigmoid + sigmoid function with ReLu(Rectified linear units)+ sigmoid function toextract better efficiency painting images representation.The experimental results prove that the method is effective in Chinese painting images classification task than traditional.