Chinese character field detection method and system based on character recognition

Abstract

The invention discloses a Chinese character field detection method and system based on character recognition, and the method comprises the steps: recognizing a character region in a pre-selected region, collecting the character region, and calculating a relative distance matrix; clustering the matrix based on DBSCAN to obtain a character string region; extracting characters from the character string area through a sliding window and then putting the characters into a single-character classifier, and obtaining a predicted Chinese character field; training a single character classifier forwardsand backwards through CTCLoss on the basis of the predicted Chinese character field, and outputting the character probability through a softmax function; and putting the characters extracted by the sliding window into the trained single-character classifier to obtain a Chinese character field. Through the relative distance matrix word clustering and forward and reverse CTCLoss training of a sliding window classifier, Chinese character fields in a complex environment can be accurately identified. The problems of inaccurate character string extraction and overlarge model time and space complexity are solved, and the Chinese character field in a complex environment can be identified more accurately.

Description

technical field [0001] The invention relates to the technical field of character detection, in particular to a method and system for detecting Chinese fields based on character recognition. Background technique [0002] OCR (Optical Character Recognition, Optical Character Recognition) technology can scan documents quickly and accurately, but the traditional OCR solution is to use edge detection technology to obtain text boxes, and then use font file matching for recognition, which can only recognize documents with high resolution and Images with simple backgrounds cannot recognize complex backgrounds and text with environmental noise. Therefore, traditional image preprocessing methods rely on standard image data. [0003] The emerging OCR scheme uses the method of deep learning, through the recognition method of Two-stage (locating the character string box, and recognizing the text in the box). Existing positioning methods in natural scenes include: CTPN (Connectionist Te...

AI Technical Summary

Problems solved by technology

However, the CTPN positioning method for the positioning frame cannot locate oblique or irregular character strings; the Seglink positioning method needs to specify the positioning of each character and its inclination angle θ, which requires high quality training data; The positioning and cutting method of EAST has relatively high requirements for training labels, and training labels for pixel sampling are required

[0004] The long text recognition means in the prior art has CRNN (Convolutional Recurrent Neural Network, Convolutional Recurrent Neural Network) and Attention (Attention Network), both of which use CTCLoss (Connectionist Temporal Classification) as the loss function. Neural network) method, which has low computational efficiency and is easy to overfit; the latter Attention performs self-attention on the features of candidate boxes, requires a large storage space for calculation, and is not easy to deploy on a single machine

Images