Method and device for generating text line classifier

Abstract

The present application discloses a method and device for generating a text line classifier, wherein the method for generating a text line classifier includes: using the current terminal system font library to generate text line samples; performing feature extraction on the text line samples and pre-stored marked samples; And perform model training according to the extracted features to generate a text line classifier for identifying text regions. The generation method of the above text line classifier is based on the method of generating text line samples in the system font, so that the generated text line classifier can identify text regions for different scenarios or different needs, has strong applicability, wide application range and simple implementation, and at the same time The feature extraction method of text line samples combined with labeled samples makes the generated text line classifier have high accuracy.

Description

technical field [0001] The present application relates to the technical field of pattern recognition, in particular to a method and device for generating a text line classifier. Background technique [0002] At present, many pictures such as Taobao pictures contain a large number of prohibited texts. In order to identify these prohibited texts, the optical character recognition (Optical Character Recognition, OCR) technology of natural scene pictures can be used to filter the results of text detection and positioning, and filter out non-text The detection results are selected, and candidate texts are screened out and sent to the recognition device, thereby improving the accuracy of recognition. [0003] Among them, the OCR technology of natural scenes has always been one of the hotspots in the industry and academic research. For different languages, the features used and the algorithm architecture will change. At present, the OCR technology in the world is mainly for Englis...

AI Technical Summary

Problems solved by technology

[0005] In the existing text area recognition of Chinese OCR, the method of classification using empirical threshold is the simplest, and the features for judgment are mostly from the text features extracted from single-character verification, but the algorithm has low accuracy and robustness. Poor, prone to over-fitting phenomenon; the second type of method is currently the mainstream solution, the use of the third type of method is rare, the main reason is that the CNN method will consume too much computing resources and affect the overall efficiency of the algorithm, but , whether it is the second type of method or the third type of method, a large number of samples need to be labeled, which will inevitably consume a lot of labor costs, and the classification effect depends on the extraction of features and the selection of samples, so it is often necessary for different application requirements. Annotate a batch of new business-dependent data, that is, new samples. Therefore, the existing annotation samples have poor applicability. Not only that, Chinese characters have various fonts and complex styles, including simplified, traditional, and handwriting. The diversity is extremely rich, which undoubtedly greatly increases the difficulty of identifying Chinese text areas.

Images