Natural scene text recognition method and system of multi-path parallel position association network

Abstract

The invention relates to a natural scene text recognition method and system for a multi-path parallel location-associated network. Comprising the following steps: step A, cutting an image only containing text information in a natural scene picture, labeling a corresponding text and a text length in the picture, and constructing a training data set S; step B, using the training set S to train a deep learning network model G based on multi-path parallel position association, wherein the deep learning network model G is used for identifying text information in a natural scene picture; and step C, inputting the text picture which is cut into a fixed size into the trained deep learning network model G to obtain corresponding target character text information in the picture. According to the invention, the text recognition accuracy can be effectively improved.

Description

technical field [0001] The invention relates to the field of computer vision applications, in particular to a method and system for recognizing text in natural scenes with a multi-channel parallel location-associated network. Background technique [0002] Scene Text Recognition (STR) is a basic task of text recognition tasks in computer vision, and its basic goal is to recognize character text from complex natural text images. With the rapid development of the Internet and new media, text also appears in various complex scenes, including street views, store names, advertising slogans, and product packaging, because the text descriptions in these scenes can help us understand the content of the scene objectively. information, so people have higher and higher requirements for recognizing text information in natural scenes. Therefore, researchers have proposed recognition tasks based on text images in natural scenes. Existing scene text recognition models are mainly divided i...

AI Technical Summary

Problems solved by technology

[0007] Most of the current models are less robust, and the various shapes and curved patterns of irregular text cause greater difficulty in recognition

On the one hand, due to the complex image background, the adjacent characters are closely glued together, which is prone to recognition errors

On the other hand, mainstream recognition networks only consider local sequence context dependencies. When predicting character sequences, they lack the supervision of global semantic information and will misidentify edge features.

Images