Text recognition method, device, device and storage medium

Abstract

The present disclosure relates to a text recognition method, device, equipment and storage medium. The method includes inputting the obtained preset object image into a progressive expansion network for segmentation to obtain target segmented images including irregular text regions; using a thin-plate spline interpolation algorithm to correct the target object image, and using a text recognition model to identify the corrected Horizontal object text image. The pixel-level segmentation of preset object images can be realized to effectively detect text areas of various shapes, and by correcting irregular text areas to obtain horizontal text areas for recognition, it is possible to avoid text errors caused by direct recognition of irregular text Out-of-order problems, as well as a large amount of labeling work for irregular text, text recognition models that can be trained with horizontal text images have stronger generalization capabilities. In addition, the correction is specifically implemented through the TPS algorithm, which can be applied to more complex application scenarios and has better robustness.

Description

technical field [0001] The present disclosure relates to the technical field of artificial intelligence, and in particular, to a text recognition method, apparatus, device and storage medium. Background technique [0002] At present, OCR (Optical Character Recognition, Optical Character Recognition) technology is widely used. When OCR technology detects text, the detected text area is generally a rectangular frame, a rotated rectangular frame or a four-point rectangular frame, such as figure 1 As shown, in this detection method, for irregular text (such as curved text), the detected text area will include a large number of background areas, which will cause great interference to the recognition of the text. Moreover, the existing text recognition methods, such as the Convolutional Recurrent Neural Network (CRNN) recognition method, are only effective in recognizing rectangular text images, but cannot recognize text images including a large number of backgrounds. [0003] In...

AI Technical Summary

Problems solved by technology

[0002] At present, OCR (Optical Character Recognition, optical character recognition) technology is widely used. When OCR technology detects text, the detected text area is generally a rectangular frame, a rotated rectangular frame or a four-point rectangular frame, such as figure 1 As shown, this detection method, for irregular text (such as curved text), the detected text area will include a large number of background areas, which will cause great interference to the text recognition

Moreover, the existing text recognition methods, such as the convolutional recurrent neural network CRNN (Convolutional Recurrent Neural Network) recognition method, are only good for rectangular text image recognition, but cannot be recognized for text images that include a large number of backgrounds.

[0003] In addition, the existing recognition methods for irregular text include attention model (Attention model) and polar coordinate correction method, wherein the attention model can recognize the text of 2D structural information (such as formula), and can be applied to the irregular text Recognition, but requires a large amount of training data, and it is difficult to construct through simulated data. At the same time, it will introduce the problem of chaotic text characters

The polar coordinate correction method is to restore the arc-shaped text to a straight-line text and then perform text recognition. This method is not robust to lighting, distortion and complex scenes.

[0004] Nowadays, in the cloud operations of many online businesses, the text of certain types of objects in the image will be recognized and authenticated, such as the text of the seal in the images of official documents, bills, and certificates, etc., and the seal will generally include Irregular text (such as curved text, T-shaped text, etc.), from the above analysis, it can be seen that none of the existing text recognition technologies can effectively identify the irregular text in the seal

Images