Natural Scene Text Detection And Recognition Based On Deep Learning

Text detection and recognition technology in natural scenes is widely applyed in fields such as unmanned driving and smart transportation.Compared with traditional document recognition,the text background in natural scenes is complex and scattered.It makes detection more difficult.In order to further improve the accuracy and speed of detection and recognition,this thesis proposes an end-to-end text detection and recognition method based on deep learning algorithm,which uses a combined structure of convolutional neural network and bidirectional recurrent neural network as a feature extraction network to fully extract text Sequence characteristics.Through the shared feature extraction network and the joint training loss function,the detection and recognition modules are closely combined and then the text positioning and category information are directly obtained at the output end.The research content of the thesis is divided into two parts: text detection and text recognition:(1)Text detection adopts a method based on the idea of candidate boxes.Combining the distribution characteristics of the text and analysising of the performance of several conventional target detection algorithms,this thesis finally chose to make better based on YOLOv3 by improving the nerwork structure and adjusting the loss function.In order to realize the detection of text at any angle,the thesis adds the angle information of the text box during the training process.Candidate boxes are screened by using the combined judgment method of area intersection ratio and coincidence degree to further improve the accuracy of text box bounding box regression.(2)Text recognition method is based on sequence and the network utilizes the encoding-decoding structure of fusion attention mechanism.The structure can make up for the shortcomings of the poor encoding and decoding structure for long text sequence recognition.So that the context information of the text sequence can be better exploited.The encoding network uses the feature extraction network of the detection part and the decoding network uses the conventional LSTM network structure.After training on multiple standard data sets and adjusting network parameters,the final experimental results prove that the improved algorithm network in this thesis has excellent performance on multiple evaluation standards.The detection speed can reach 16 frames per second and the highest recognition accuracy rate is 86.47%.The comparison with the result data of related classical methods further shows that the improved algorithm network in this thesis is more reasonable and effective.