Universal detection algorithm for optical printing text and scene text

Abstract

The invention discloses a general detection algorithm for an optical printing text and a scene text, and the algorithm comprises the following steps: S1, obtaining a to-be-detected picture, S2, detecting a target text region in the to-be-detected picture, S3, determining a detection model of a target text corresponding to the target text region according to a preset corresponding relation between a text region and a text detection model, and S4, obtaining a target text detection box by adopting the target text detection model. In the invention, the DBnet network algorithm model can test the performance of the detection algorithm in the text recognition under the condition of not changing the source code of the text recognition model, and adjust the text detection algorithm based on the test result, so that the accuracy of the text detection algorithm can be improved, the accuracy of the whole text recognition model is also improved, the text detection integrity can be improved, and the problem that in the prior art, various text types in a complex layout cannot be completely detected by adopting a single process is solved.

Description

technical field [0001] The invention relates to the technical field of text detection algorithms, in particular to a general detection algorithm for optically printed text and scene text. Background technique [0002] OCR (Optical Character Recognition, optical character recognition) is to convert the text of various documents, bills, newspapers, books, manuscripts and other printed materials into image information through optical input methods such as scanning, and then use text recognition technology to convert image information into images that can the computer input technology used; [0003] It plays a vital role in applications such as image search, instant translation and robot navigation: for example, identifying signs, street signs, store names, etc. in photos and videos, OCR is good at recognizing black and white characters on planes, and text in natural scenes has Complex lighting, occlusion, angles, fonts and 3D shadows, compared to the mature document text recog...

AI Technical Summary

Problems solved by technology

[0005] First of all, the existing technology is relatively mature for text detection in a single direction, such as horizontal and vertical directions, but it cannot achieve an accurate detection of scene texts of various shapes (curved, vertical, multi-directional), based on bank For bill text recognition, in some business scenarios, it is necessary to detect the seal in the bill, so for bank bill text recognition, text detection algorithms for various shapes are necessary;

[0006] Secondly, most of the text in natural scenes are scene texts of various shapes. The existing technology uses segmentation methods, but usually the post-processing process is complicated, time-consuming, sensitive to thresholds, and affects the effect.

Images