Calligraphy Style Recognition Based On CNN

Chinese calligraphy is one of the treasures of the splendid culture of the motherland.With the development of technology and the improvement of the quality of life,the state's investment in cultural undertakings and the public's pursuit of literature and art are showing a trend of continuous improvement.More and more artworks are collected and archived,and electronic documents are made for enthusiasts.There are many different styles of calligraphy fonts.Usually beginners learn to copy a specific style.However,beginners cannot accurately determine the type of a particular font.Therefore,it is important to correctly classify the collected calligraphy samples.The practical significance.However,the computer is not perfect in automatically recognizing the style of calligraphic fonts.Traditional machine learning methods can classify samples.However,the premise is that the feature extraction method is used to extract features from the samples.Because the features of the calligraphy font sample are artificially defined,The extracted features have a certain degree of subjectivity,and the classification results are classified hard,resulting in a high rate of misclassification.The machine learning methods can not automatically extract image features,and can only do simple classification processing on the extracted features.This article under the framework of the Convolutional Neural Network(CNN),based on the Maximum Pooling(Max Pooling)and Average Pooling(Avg Pooling),we put forward the Local Pixel Probability Normalization Pooling(LPN Pooling).The LPN method combines the principle of large value decision of the Max method and the principle of equal distribution of the Avg method.Themethod first calculates the sum of all the pixels' values in the pooling area,and then calculates the weights of the respective pixel points in the sampling area.The weights are taken as Each pixel participates in the degree of contribution of the sample.Each weight value is multiplied by its corresponding pixel value.The cumulative summation of all products is used to obtain the sampling result of the LPN method.To verify the LPN method,the experimental design process is as follows:First of all,for the particularity of the calligraphy dataset,in order to carefully extract the local characteristics of the sample,it is necessary to determine the best sampled kernel specification.Since the classical data set CIFAR-10 contains many common objects,it is suitable for the research of multi-classification problems,and all the samples are in RGB format,covering more pixel areas and more suitable for experimental verification than binary data sets.Customize the design of the network structure and embed the three sampling algorithms of Max,Avg,and LPN in the same network structure.Finally,set up the samples with different specifications to compare the experiments,and obtain the three networks when the sampling core is 3×3.Have achieved the best classification results.Secondly,the three networks are trained on multiple classical data sets.The performance of each network is evaluated based on the comprehensive performance of training accuracy and test accuracy.From the experiment,it can be concluded that the training performance and test performance of the LPN method have the smallest gap.The LPN network has the best average accuracy.Therefore,the LPN network is selected as the optimal network model.Finally,experiments on calligraphic dataset are carried out using LPN networks and various machine learning classification methods.From the experimental results,it can be clearly seen that the rigid division error rate is the highest,the fuzzy classification is better,and the BP neural network has agreater improvement.The highest classification accuracy is obtained by the CNN.Theoretical verification and experimental results show that the results of the LPN pooling method are between the MAX pooling method and the AVG pooling method.It is confirmed that the principle of the LPN to participate in the decision-making of all pixels is the working principle,and the sampling method can be further improved.The smooth and balanced feature extraction effect,because the feature image is sufficient and objective,the classification accuracy is obviously superior to the Max and Avg sampling methods.At the same time,it is verified that the CNN has higher recognition accuracy than the traditional classification algorithm because of its ability of self-extraction and self-learning.Therefore,this article has certain reference value in the recognition of calligraphy style.