A Research Of Natural English Character Recognition Based On Deep Learning

As a common form of information in natural scenes,text often contains rich semantic feature information.This thesis studies the problem of English text recognition in natural scenes.Different from the document image used in the traditional OCR method,the text images collected in the natural scene often have difficulties such as variable fonts,irregular layout and complex environmental background,which introduces many uncertain factors for the recognition process.How to make computers more accurately extract features and recognize words has always been a hot issue for many scholars at home and abroad.With the continuous development and advancement of deep learning theory,its application fields have gradually expanded,and achieved good results in many directions such as image processing and forecast classification.Based on the deep learning theory,this thesis aims at the recognition of English characters in natural scenes and uses the accuracy and efficiency of text recognition as the standard.The main contents are as follows:(1)Combined with the idea of end-to-end recognition model,a set of English text recognition algorithm model based on CNN-BiGRU-Attention is designed in the framework of encoder-decoder.The recognition task is decomposed into two processes of coding and decoding.And then this thesis carries out research and design on each part of the algorithm.(2)In the coding process,on the basis of image preprocessing,combined with Jaderberg's convolutional neural network,a CNN model for image static feature extraction is presented.In order to further utilize the contextual information,a bidirectional recurrent neural network is introduced to extract the associated information.At the same time,in order to further reduce the computational complexity,a gated recurrent unit is adopted as its basic structure.Based on the extraction of static feature and associated feature,the context feature vector generation method is proposed,which realizes the feature coding of the original image.(3)In the decoding process,the recurrent neural network is used to decode according to the context feature vector.The attention mechanism is introduced here to enhance the application of context feature information,thereby improving the overall accuracy.Combined with the characteristics of the research data,the thesis optimizes the attention mechanism,proposes the local attention mechanism used in the decoding process,enhances the application of local correlation features,and reduces the overall calculation.In the post-processing stage,the cluster search algorithm is optimized by two methods based on the lexicon model and the n-gram language model,which further improves the accuracy of the decoding result.(4)In the process of model training,the methods of exponential decay,regularization and moving average model are introduced,and the algorithm model is further optimized from two aspects of efficiency and recognition accuracy.After finishing the model training,the efficiency,accuracy and generalization ability of the proposed algorithm are tested from several aspects.