Research On Text Recognition And Emotion Analysis Based On Deep Learning

Text,as a unique symbol of human beings,contains high-level semantic information and rich emotional information.If the computer can accurately recognize and understand the content of text information,it has very important engineering and academic significance.Text recognition in complex scenes has been widely used,such as license plate recognition,virtual reality,human-computer interaction and industrial automation.However,many texts,such as those in natural scenes,are often accompanied by a series of problems such as complex image background,low text resolution,smudged text images,diversified text fonts and irregular distribution.However,such words contain a lot of valuable information,from which a lot of commercial and human value can be obtained.Therefore,it is worth exploring and studying how to accurately locate and identify characters in natural scenes.In this paper,text positioning and recognition of natural scenes are studied in combination with the rapidly developing deep learning technology.As an extension of text recognition,emotional analysis of the recognized text is also carried out to explore more information value.The main work of this paper is as follows:Compare several neural network models in deep learning,learn their structure,advantages and disadvantages,and learn how they play their roles in text positioning,text recognition and emotional analysis tasks.Through the study and research of neural network model,the localization model is improved,and the localization robustness is improved through the method of single character localization and relinking to generate candidate box,so as to reduce the structure of the model and speed up the training of the model.The localization content is then identified by convolution extraction feature and CTC to complete the task of text localization and recognition.The performance of fast Text model with simple structure and BiLSTM+Attention model with relatively complex structure in emotional analysis task were compared,as well as the performance of the two models with different quality word vectors.