Text Detection And Recognition Based On Deep Learning

Image is an important carrier which contains a large amount of information such as text information.The extraction of text information can promote the understanding of images.Compared with the text information extraction of a specific scene,the text detection in the natural scene is more difficult.Due to the folding and curling phenomenon of documents such as newspapers and bills,it is hard to extract the text information from document images.To solve these problems,this paper proposed a model for document image correction and a model for text information extraction.The specific work includes the following contents:1.The Stacked-EAST model is proposed to correct distorted document images.The model predicts the offset data for each pixel of the distorted document image to correct the image,and the training data can be self-generated from the flat document image,which is easy to understand and has a wide application range;2.A text detection and recognition model suitable for natural scenes is proposed.This paper combines the two-stage target detection model with image segmentation,and solves the problem that the long text is difficult to be covered.The maximum merged non-maximum suppression algorithm is proposed to improve the text in the document environment such as bills.After all,CRNN model is used to recognize text.The text detection and recognition model proposed in this paper has achieved excellent results on multiple datasets,and achieved the 10 th place in the end-to-end network image text detection and recognition challenge of the ICPR MTWI 2018,which proved its effectiveness.