A character recognition and processing method for image acquisition by a digital screen

Abstract

The invention discloses a character recognition processing method for images collected by a digital screen. The characteristic map output by a character detection model consists of two values, the text area and the background area, respectively, and the text area is extracted from the image collected by the digital screen through a mask. ; The input image of the text recognition model is a preset size, and the output of the text recognition model is a character category, and the character area text area is obtained by querying the mapping relationship between the category value and the computer character; aggregation is performed according to the row coordinates of the character area text area, and According to the sequence of each character row coordinates from left to right, the characters are composed into a string; the training phase of the text detection model introduces text edge and text gap images as training data, and in the inference phase, when the center of the sampling window falls on the text edge or between the two When there is a space in the middle of a text, the final feature value is defined as the background. The present invention can simulate the text features of the text screen and generate characteristic training data, so that the model can realize targeted training on the text of the text screen, and the training effect is good.

Description

technical field [0001] The invention relates to the technical field of word processing, in particular to a method for character recognition and processing for image acquisition by a digital screen. Background technique [0002] At present, text recognition (OCR) in images is an important scene in the field of deep learning. Different from the traditional method of using image processing technology to extract text area features and using classifiers to determine characters, OCR technology based on deep learning uses deep neural networks to extract image features, which can achieve much higher recognition accuracy than traditional effects. [0003] OCR processing based on deep learning is generally divided into two models, text area detection and text character recognition. The text area detection model scans the input text image and marks the text area; the text character recognition model extracts and classifies each character in the text area to obtain the character value,...

AI Technical Summary

Problems solved by technology

[0006] First, compared with the printed text, the captured image of the digital screen has uneven, jagged, and often distorted edges. The text recognition algorithm for electronic documents is prone to misrecognition when dealing with the above situations;

[0007] Second, the general text detection model and text recognition model are complex and the training cost is high;

[0008] Third, the text and image recognition of the digital screen generally needs to be deployed on-site embedded devices, and real-time recognition is required

However, this method has poor generalization for text segmentation and cannot cope with the situation where the text spacing is too close or there is noise.

Images