Scene text recognition method and system based on attention mechanism

Abstract

The invention discloses a scene text recognition method and system based on an attention mechanism, and the method comprises the steps: obtaining a scene text image data set, and carrying out the preprocessing; constructing a scene text recognition model, and inputting the preprocessed scene text image data set into the scene text recognition model for model training; the scene text recognition model comprises a conversion network used for correcting a scene text image into a rule, a feature extraction module used for extracting the corrected scene text image as a global visual feature, and a multi-layer attention decoder; and obtaining a test set of a to-be-detected scene text image, inputting the test set into the model, and obtaining an identification accuracy rate as a performance evaluation index of the model for evaluation and detection. According to the method, the improved scene text recognition model is constructed, a global attention mechanism and a self-attention mechanism are introduced into the model, and a multi-layer superposed attention decoder is fused, so that the decoding capability of the model on global visual features and global sequence features is effectively improved.

Description

technical field [0001] The invention relates to the technical field of computer vision, in particular to a method and system for scene text recognition based on an attention mechanism. Background technique [0002] In recent years, with the development of deep learning, some progress has been made in text detection and recognition methods based on deep learning, and different feature extraction networks have also been applied to scene text recognition. Scene text recognition models can be divided into three categories from the perspective of text targets: the first category is a recognition model based on word classification, which requires a word dictionary to be designed first, and each word in the dictionary is regarded as a category in the classification task, so that The task of converting text recognition into image classification; the second category is a text recognition model based on character recognition, which usually needs to be matched with a character-level te...

AI Technical Summary

Problems solved by technology

[0004] The disadvantage of the existing technology is that the text recognition method has insufficient ability to deal with complex scenes, and cannot effectively extract the visual features of text objects, resulting in low recognition performance for text images in irregular scenes

In addition, in the existing text recognition models, the feature extraction module mostly uses a general image classification task model, and the feature extraction ability for text objects is insufficient.

Finally, the existing methods are not capable of modeling the dependencies between text feature sequences, and cannot effectively model long-term dependencies, and the training of the modeling part takes a long time and is difficult to converge

Images