Research And Implementation Of Scene Text Image Multi Classification Based On Deep Learning

With the outbreak of the information age,the Internet is flooded with heterogeneous data.As the main data type,text and image data contain rich information and patterns.As the organic combination of the two types of data,text images in the natural scene provide important source of information for our society,and can help us to use various types of applications,including information retrieval,human-computer interaction,driving navigation and other fields.Therefore,the field of text-image related research has also been a topic of great concern.The problem of multi classification of natural scene text images has an extremely important application requirement in real life,but few people directly study the problem in this field.It is usually a separate task of text classification or image classification task,and it does not effectively combine the relevance of text content and image scene in text image.This paper is the first time to study and discuss the problem in depth,so as to promote the recognition and classification of the text image field,and to accelerate the application of the task in the actual scene.Therefore,this paper will focus on the work of text image multi classification.The research contents and innovations include the following points:1)First,this paper applies text image generation algorithm to generate the natural scene text image dataset by capturing the network image and corresponding text data,which is used for the dataset of following chapters.2)In order to improve the accuracy of text image detection and localization,this paper proposes a text region recognition algorithm based on Multi Scale Connectionist Text Proposal Network,which extracts the candidate text frame for the text image classification task,and compares the performance comparison between the generated data set and the other text image recognition methods;3)For the sake of improving the accuracy in the process of text extraction,this paper proposes an end to end text sequence extraction method based on attention mechanism,which translates text content directly into the corresponding text sequence and effectively improves the accuracy of recognition and extraction;4)For extracting the inherent pattern of the natural scene text image,this paper proposes a joint multi classification network based on deep features in text and image.This multi classification network combines the patterns contained in the image and text,and integrates the image features learned by the Convolution Neural Network and the character of the text sequence learned by the Recurrent Neural Network.