Research On The Key Technonlogies Of OCR For Real Scene Text

With the development of society,artificial intelligence technology affects people’s life to a certain extent,making life more convenient and work more efficient.In our daily life,people’s demand for the collection and processing of text information in real scenes is increasing day by day.How to process and obtain a large amount of scene information has become the focus of attention,and also become the focus of academic research.In many of the real scene,first of all,we focus on paper,based on indepth study of the paper content recognition and support more than 8000 characters,we select PSENet paper text detection algorithm,selecting CRNN text recognition algorithm for paper work,and making the dataset,to identify the main job is to improve the identification of papers.After that,the accuracy of text detection and recognition is improved by optimizing,adjusting parameters and preprocessing.At the same time,by collecting new characters,corpus and existing paper pictures to make a dataset,plus the existing datasets,completing paper recognition support 8000 characters set,and improving the recognition rate of paper through training model and meeting the needs.The corpus includes new character corpus,name and address corpus and paper specific field corpus.The production of dataset also left some reflections.First,for the training model,the training data samples provided should not only have good and clear quality,but also need low quality data samples to improve the model generalization ability.However,the existing datasets do not contain low-quality samples,so it is worth considering how to generate low-quality samples of real scenes,so it is a research point of this paper.Under normal circumstances,the range of low quality refers to fuzzy,partial distortion,but does not affect the overall information acquisition;Second,from the result of paper recognition,the low accuracy of the detection of deformed text is an important reason for the accuracy of paper recognition.Similarly,we also take it as another research point.Therefore,in addition to making the dataset,this paper also focuses on the generation of low-quality samples of real scenes and deformation text detection of real scenes.In addition to the dataset,the above two problems should be solved.First,in order to generate low-quality samples,I proposed a new generator and introduced noise features to generate low-quality pictures,and conducted training and testing on the dataset.From the results obtained,the function of generating low-quality samples was realized.Second,for bent or deformed text detection in real scenes,we adopt the method of optimizing the existing algorithm TextSnake to solve the problem.At present,we use two ideas to optimize the network model of the algorithm.The first is the direction of graph neural network and the second is the cascade direction of network model.In the end,the latter performs better than the former in terms of related evaluation indexes of deformation text detection,and improves by one or two percentage points compared with TextSnake algorithm.