Research On Scene Text Detection And Recognition Algorithm Based On Deep Learning

Text is the cornerstone of effective communication of human civilization.By recognizing complex images and texts in natural scenes,a large amount of information can be obtained and transmitted.The scene text detection and recognition technology has great significance and wide application in many fields such as document retrieval,certificate recognition,license plate recognition,robot vision and so on.In view of the problem of information loss of different degrees in arbitrary shape text images in natural scenes due to illumination,scale and difficulty in distinguishing text lines,this paper proposes a scene text detection and recognition algorithm based on deep learning.The main research contents are as follows:(1)This paper firstly introduces and briefly analyzes the domestic and foreign classical algorithms in the field of text detection and recognition in recent years.On this basis,segmentation-based detection algorithm is selected for deep learning.Finally,ANN,CNN,RNN and Res Net are taken as examples to illustrate the development process of neural network,and the datasets and related indicators used in the following paper are introduced.(2)In this paper,Res Net50,a deep residual network based on convolutional neural network,is used as the backbone network of the text detection model.Then,using the idea of ASPP for reference,instance regularization and parallel void convolution module are introduced into the pyramid network structure with increased cross-layer connections to extract more semantic information.In addition,this paper adopts the method of multi-scale feature fusion to learn the features of different scales,so as to predict the text instances of different kernel sizes.At present,segmentation-based methods are mostly used to detect image text in any direction,so PSE method is chosen to gradually expand the text line area,so that the model can obtain detection results of image text in any direction.(3)In this paper,the attention mechanism is introduced into the classic CRNN recognition model,and the VGG-16 feature extraction network is improved,so that the recognition model can realize more efficient text recognition of any length scene.Aiming at the problem of unclear semantics caused by the inability of traditional recognition algorithms to segment long texts,the Viterbi word segmentation algorithm is introduced in this paper,and through the combination with the detection model,a complete scene text detection and recognition framework is formed.The text detection algorithm based on PSE in this paper was tested on SCUT-CTW1500 and ICDAR2015 datasets of natural scenes respectively,the results show that the proposed algorithm in accuracy and recall rate and F1 on the value of three indicators reached 88.5%,77% and 81.3%,respectively,on the premise of without adding a pre-training module,the detection effect of this algorithm on curved text is better than other detection text;The text recognition algorithm in this paper was tested on ICDAR2013 datasets of natural scenes,and the results showed that the recognition accuracy reached 94.5%.Compared with the original CRNN algorithm,the recognition effect of long text was better and the readability was better.