Research On Painting Image Classification Based On Convolutional Neural Network

Painting is an important vehicle for the development of human civilization,the study of painting can help people better understand human history and culture,it is significant to realize the digitalization of painting works for the effective use of painting resources.The methods of traditional image classification do not take into account of the subjective characteristics of the painting works,and most of the features need to be extracted manually,thus it may cause problems of easily missing detailed features and the insufficient generalization ability of the model.In recent years,convolutional neural network has been outstanding in the field of computer vision.By imitating the cognitive of human brain,convolutional neural network can automatically learn characteristics of image data,it not only avoids the artificial extraction of complex features,but also improves the accuracy of image classification and recognition.In this paper,a method of painting image classification based on convolutional neural network is proposed to improve the accuracy of classification of painting images.Although the convolutional neural network is excellent in the field of image classification,it is difficult to prove its solution process by mathematical theory,thus the network structure and parameters are optimized mainly by experiment,so as to improve the classification performance of the model.In addition to studying the convolutional neural network itself,the combination of convolutional neural network and other machine learning methods,introducing the advantages of other methods to solve corresponding problems in specific tasks is also a hot research topic.In addition to studying the design and optimization of the convolutional neural network itself,this paper also introduces the fusion of transfer learning and convolutional neural network model to solve the classification problem of small sample painting image,at the same time,introduces the fusion of weak supervised learning and convolutional neural network model to make a full use of a large number of unlabeled samples.the main contents are as follows:(1)Research on the design and optimization of convolutional neural network inthe classification of painting images,according to two painting images datasets of the east and the west,the influence of the structural depth of convolutional neural network on the accuracy of classification model is studied,so as to set the appropriate depth for convolutional neural network structure.In the structure of the traditional series convolutional neural network,the influence of the size of convolution kernel on the accuracy of classification model is studied,so as to set different sizes of convolution kernel for different datasets.At the same time,in the basis structure of the traditional series convolutional neural network,the width of the network structure is increased,and the multi-size feature extraction and fusion is realized,which enhances the feature learning ability of the convolutional neural network.In addition,the influence of the number of training samples on the classification results is also analyzed,it is found that the increase of the number of training samples can enhance the learning ability of the convolutional neural network.The experimental results show that the proposed method has good classification performances for painting images.(2)Research on convolutional neural network in the classification of small sample painting images,for small sample painting image dataset,the number of training samples is first amplified by data expansion method such as slice,rotation and mirroring to enrich the sample diversity.Then a new activation function PPReLU is proposed to reduce the variance of eigenvalue by introducing the learning parameters to accelerate the convergence and prevent the occurrence of gradient explosion.Last,the transfer learning is introduced into the convolutional neural network,and a combination transfer learning method is proposed,the weights of the pre-training model are transferred to the painting image classification model in two stages.The experimental results show that the proposed method can effectively improve the classification accuracy for small sample painting image dataset.(3)Research on convolutional neural network in the classification of weak supervised painting images,we use painting image classification model trained from labeled samples to automatically label a large number of unlabeled samples,so that we can make a full use of a large number of unlabeled samples for weak supervised learning.Then we use hard sample mining method to study hard samples,experiments show that adding small proportion of hard samples can improve theclassification performance of the classification model on the test dataset.Finally,an improved softmax loss function is proposed,through automatic learning weight adjustment factor,the weights of simple samples are automatically reduced during training stage,and the weights of hard samples are automatically increased,so that the classification network can better converge hard samples after the convergence of the simple samples in the training stage.The experimental results show that the proposed weak supervised learning method has good classification performances on painting images,so our method is feasible and practical.