Natural scene character detection and recognition method based on CRAFT and SCRN-SEED frameworks

Abstract

The invention discloses a natural scene character detection and recognition method based on CRAFT and SCRN-SEED frameworks, and the method comprises the following steps: (1) building an image data set through employing a real data set and a synthetic data set, and dividing the image data set into a training set and a test set; (2) training a CRAFT network by using the image data set; (3) training an irregular text correction network SCRN by using the real data set; (4) combining the SCRN with the SEED network, and training the combined SCRN-SEED network; and (5) the CRAFT network is connected with the SCRN-SEED network, and a complete model is constructed and trained. According to the method, bent and deformed texts or long text instances can be fully detected, each character is accurately positioned, then the detected characters are connected into one text through an affinity mechanism to achieve the purpose of detection, and the method is suitable for bent, deformed or extremely long texts; by correcting irregular text pictures and using semantic information for global information detection, low-quality text instances can be accurately recognized.

Description

technical field [0001] The invention relates to the technical field of text detection methods, in particular to a natural scene text detection and recognition method based on CRAFT and SCRN-SEED frameworks. Background technique [0002] Optical Character Recognition (OCR) traditionally refers to the analysis and processing of the input scanned document image to identify the text information in the image. This type of technology assumes that the input image has a clean background, simple fonts, and neatly arranged text. Under the circumstances, a higher recognition level can be achieved. Scene Text Recognition (STR) refers to the recognition of text information in natural scene pictures, which is much more difficult than text recognition in scanned document images. Natural scene text display forms are extremely rich, and there may be the following situations: Horizontal, vertical, curved, rotated, twisted and other styles; the text area in the image is incomplete, blurred, e...

AI Technical Summary

Problems solved by technology

[0003] The scene text detection method based on the neural network has achieved good results, far surpassing the traditional technology in detection and recognition, but it still cannot solve the problems of bending, blurring and interfering textures in natural scenes.

The existing text detection methods in natural scenes have the following problems: the text positioning is not accurate, and the traditional text positioning model framework mostly focuses on the entire line of text, requiring a large receptive field, and using a single rectangular frame to mark the text location, which is very important for Curved, deformed or extremely long text is not suitable, and it is difficult to accurately locate the label; text recognition is not accurate, and many recognition methods based on the encoder-decoder framework are proposed to deal with curved text, but most of them are based on local visual features. However, the global semantic information is ignored, so the recognition accuracy is greatly reduced in the face of situations such as blurred images, uneven lighting, and incomplete characters.

Images