Research On Calligraphy Style Recognition Based On Convolutional Neural Network

Calligraphy is an important part of the Chinese traditional culture and art.Using the methods based on image processing and pattern recognition for calligraphy recognition,retrieval and style appreciation is of great significance.In recent years,deep learning technology has been applied more and more deeply and widely in the field of image and recognition.Aiming at the shortcomings of traditional image feature descriptors in image representation and the disadvantages of low recognition rate,this paper proposes a calligraphic style recognition method based on deep learning,applying convolutional neural networks to calligraphic style recognition,and greatly improving the accuracy of style recognition.The main work of this article includes the following three aspects:First of all,a systematic analysis of the status quo of domestic and foreign research on the recognition of calligraphy styles is carried out,and the text detection,recognition and recognition of calligraphy styles and the content of this article are elaborated.Based on this,the traditional image feature extraction and recognition methods are summarized.In terms of feature extraction,Gabor,wavelet,and Gist features are studied.In terms of recognition methods,SVM classifiers and optimization strategies are studied.Then,this paper proposes a calligraphy style recognition method based on convolutional neural network.The recognition methods based on LeNet-5,GoogleNet and ResNet,the network structure and training methods of these three models are studied.In the TensorFlow frame,these models are applied to the recognition of calligraphy style.On the standard Windows fonts and the four calligraphy datasets constructed by the research group good results are achieved.Third,the LeNet-5 model was improved,the first three layers of the convolved layer in the improved model are followed by the maxpooling layer,the fourth convolutional layer is followed by the avgpooling layer,and the BN layer is added after each pooling layer,and the ReLU function is used to replace the original function,the recognition rate has been improved on the standard Windows fonts and the four calligraphy datasets.Then the model is further improved by adding the Concat(avg,std)layer before the fully connected layer,constructing a new model named as C-LeNet(Concat(avg,std)LeNet).The model has fewer parameter parameters than the GoogleNet and ResNet models and thus the training time is shorter and the recognition rates on the Windows fonts and four calligraphy datasets of the standard script are higher than the recognition rates of the three classical models,which has reached the expected goal of the recognition of ancient calligraphy styles.Overall,the methods based on the classical CNN model and the network model constructed in this paper can more effectively abstract and transfer image features in a hierarchically progressive manner.Combined with the improvement of the model structure,the performance of the network model is greatly improved and there are good application prospects.