Research On Text Detection And Recognition In Complex Natural Scene Image

Text information in natural scene image is of great value to our daily life and provides important clues for people or computer to understand life scenes.The text information can be widely used in a variety of situations that the text in the scene image is required to be analyzed and understood,such as intelligent transportation assistance system,travel abroad automatic translation system and so on.It has been a hot research topic that computer technology is used to extract text information from these scene images automatically.Due to the complexity of the scene images,the detection and recognition of the text information need to go through three steps: First,locate the text in the image,then separate the text from background,and finally to identify.In this paper,the localization algorithm is mainly studied,and the subsequent text segmentation and recognition are also explored respectively.In this paper,a two-level localization algorithm based on multi-feature fusion and SVM classifier is proposed after studying the existing localization methods.Firstly,the scene image is preprocessed by the improved N iblack algorithm,then the morphological operations and some priori conditions are used to classify the candidate image regions roughly.Experimental results show that the proposed algorithm can remo ve a large number of non-text regions effectively,and it can provide a good foundation for subsequent localization and recognition.After analyzing the text features in the scene synthetically,the hierarchical gradient histogram PHOG feature and the loca l texture LBP feature are selected to describe the characteristics of the text character,assisted in the way of the four overall texture features.Then the candidate text regions after rough location are further positioned in combination with the SVM classifier model.The experimental results show that the combined feature is better and the accuracy of location is better.On the basis of text localizatio n,this paper studies the existing text segmentation and OCR recognition technology further.The sub-pixel technique is used to preprocess the localized area,and then the OSTU threshold segmentation method is used to segment the text effectively.A skew correction method based on least-squares fitting method is proposed for the skew text in scene image.