Research And Implementation Of End-to-end Neural Network For Text Detection And Recognition In Natural Scene

Words are ubiquitous in our daily lives and are one of the main ways of communication,information transmission and interaction between people.In recent years,natural scene optical character recognition(OCR)technology,that is,the conversion of text on a handwritten or image to machine-encoded text has become a hot research direction in the field of pattern recognition,artificial intelligence and computer vision,academia and The industry has a strong concern about this.As a general-purpose technology,natural scene text recognition does not require custom special scenes,and can recognize text in any scene image,such as billboards,road signs,license plates,document photos,merchandise packaging,and the like.The natural scene text detection and recognition technology has been widely used in information content security audit,ticket identification,document photo identification,etc.,and has extremely important research and application value.This paper discusses the research background and significance of natural scene text detection and recognition technology,expounds the research status of natural scene text detection and recognition technology,focuses on scene text detection and recognition algorithm,and designs a unified end-to-end training.The text-based deep learning network can simultaneously detect text lines and recognized text in any direction,and verifies the scene text detection and recognition algorithm on the standard data set.This article mainly does the following work:1.This paper designs a unified end-to-end deep learning network,which simultaneously completes the task of text detection and recognition.The network can be used for end-to-end training.Compared with the method of completing these two tasks by two networks respectively,the method learns more general features through the convolutional neural network,and the convolutional neural network is shared between text detection and text recognition,and the supervision of the two tasks is Complementary.Since feature extraction usually takes most of the time,it will calculate the time to shrink to a single network.2.In the aspect of scene text detection,this paper discusses the existing deep learning-based algorithm EAST,and analyzes and improves the problems of the algorithm,and solves the problem that EAST can't detect long text due to the limitation of receptive field.In order to increase the robustness of the model to more complex situations,Resnet50 was used as the base network.3.In terms of text recognition,this paper uses Convolutional Neural Network(CNN)and Recurrent Neural Network(RNN)as encoders,connected timing classifiers(The Connectionist Temporal Classification(CTC)acts as a decoder for the character recognition branch.In order to make the input sequence larger than the output sequence,the filter kernel of the pooled layer in the CNN is 2×1.4.Using the affine transformation to extract the Region of Interest(Ro I),which is the key to the combination of the detection branch and the recognition branch.The affine transformation extracts the text line in any direction detected by the detection branch from the shared feature map.Feature,sent to the identification branch for text recognition.5.Based on the research results of the detection algorithm and recognition algorithm of scene text,the algorithm is implemented based on Keras and Tensor Flow and tested and verified on multiple standard data sets.Experiments show that the scene text detection and recognition algorithm designed in this paper has better robustness.The algorithm can process natural scene images in real time,can accurately locate the position of the text in the picture,and detect and identify the text.The algorithm accuracy rate is reached.The most advanced level has strong research and application value.