Recurrent neural network-based complex image character sequence recognition system

Abstract

The present invention belongs to the image character recognition field and relates to a recurrent neural network-based complex image character sequence recognition system. The system includes an image character input module, a slide sampling module, a CNN and an RNN, wherein the image character input module is a scanner, a digital camera or an image character storage module. The slide sampling module in the system performs sliding sampling on an image character sequence to be recognized and inputs sampled sub pictures into the CNN; the CNN extracts features and outputs the features to the RNN; and the RNN recognizes the front part of a Chinese character, the back part of a Chinese character, numbers, letters or punctuations according to the input signals of the CNN, the output data of the CNN at the last time point, and vector data converted from the recognition result of the CNN at the last time point. With the system of the invention adopted, problems in the segmentation of a complex image character sequence can be solved, a language model is not required to be constructed additionally, and the recognition efficiency and accuracy of the complex image character sequence can be significantly improved.

Description

technical field [0001] The invention relates to the field of image and character recognition, in particular to a complex image and character sequence recognition system based on a recursive neural network. Background technique [0002] With the development of society, there is a large demand for the digitization of ancient books, documents, bills, business cards and other paper media. The digitization here is not limited to "photographic" using scanners or cameras, but more importantly, the Files are converted into readable and editable documents for storage. This process requires image text recognition on scanned pictures, and traditional image text recognition is optical text recognition (OCR). [0003] Conventional OCR methods include image segmentation, feature extraction, single character recognition and other processing processes, wherein image segmentation includes a large number of image preprocessing processes, such as tilt correction, background denoising, and sing...

AI Technical Summary

Problems solved by technology

[0003] Conventional OCR methods include image segmentation, feature extraction, single character recognition and other processing processes, wherein image segmentation includes a large number of image preprocessing processes, such as tilt correction, background denoising, and single character extraction; these processes The process is not only cumbersome and time-consuming, but also may cause the image to lose a lot of available information; and when the image to be recognized contains a string of multiple characters, the traditional OCR method needs to divide the original string into several small images containing a single character for separation. Recognition, and the most commonly used method for text segmentation is the projection method, that is, after binarizing the image text, find the dividing line between the two texts through vertical projection, and separate the text according to the dividing line. This method There are two main problems: 1. When the image text sequence to be recognized contains background noise, character distortion, and character bonding, it will cause difficulty in text segmentation

Especially when the image text sequence to be recognized is mixed with Chinese characters, letters, numbers, and symbols with left and right radicals, or characters in half-width and full-width formats are mixed in the image text sequence to be recognized, the character size and gaps exist due to the difference in format. The difference is that the single character in the image text sequence to be recognized cannot be accurately segmented through the simple projection method

Once there is a problem with segmentation, it is difficult to obtain accurate recognition results.

2. The method of dividing the string into sub-pictures containing a single character for separate recognition does not make full use of the dependencies between words and words in natural language. Although additional language models can be used to optimize and supplement the recognition results, but considering The construction process of language model and recognizer is independent of each other, and the optimization supplement of this method is locally limited

Images