Research On Script Identification In Text Images Based On Deep Learning

The pace of globalization of interconnection is gradually accelerating,and text images such as documents and books taken by mobile phones have become convenient and efficient forms of information exchange.Determining the language type of the text image in the text image analysis process,that is,the script identification,is an important step in the multilingual OCR technology,and it is crucial for the subsequent processing steps such as indexing and searching.However,text images are prone to affine changes and blurring distortions when photographed by mobile phones,which increases the difficulty of script identification.In order to improve the accuracy of text image script identification,this paper makes use of the advantages of deep neural network model to study script identification.The main research work is as follows:1.This paper proposes an improved convolutional neural network model for script identification.According to some shortcomings of VGG network,the last pooling layer of the network is changed to Spatial Pyramid Pooling(SPP)layer to reduce part of the semantic information lost due to changing the image size when the image is input;the image features after each convolution layer are transformed into one-dimensional vector and output to the last SPP layer,and the shallow image features are fused with the deep image features;the batch image is added after the convolution layer Normalization operation can reduce the influence of some parameters on the training results,reduce over-fitting and accelerate network training.Compared with the current mainstream deep neural network model,the improved network model has the highest accuracy,and the network effect is better.At the same time,comparing the results with the traditional text image feature extraction technology SIFT+SVM and LBP+SVM,the results are still excellent and the desired effect is achieved,which verifies the effectiveness of the improved network model algorithm.2.This paper proposes a CRNN-based text image script identification method.Combine the lightweight network model Mobile Net in the convolutional neural network with the bidirectional LSTM in the recurrent neural network.This method can directly learn the characteristics of information representation from the image data,and is not limited by the length of the sequence object.At the same time,it contains much fewer parameters than the standard deep convolutional neural network model,and it takes up less storage space.After training and testing under the text image data set,the accuracy of script identification can reach up to 97.83%.Compared with the proposed network structure,the recognition accuracy of this network model is increased by 1.91%,and the amount of training parameters is only0.46 times.Compared with the two methods of traditional text image feature extraction techniques SIFT+SVM and LBP+SVM,the experimental results are ideal.In the SIW-13 public data set,compared with the proposed network structure,the experimental accuracy of this network model is the highest.It can be concluded that the method of script identification proposed in this chapter has achieved the desired effect.