Optical character recognition method in patent text scene

Abstract

The invention belongs to the technical field of computer vision, image processing and convolutional neural networks, and particularly relates to an optical character recognition method in a patent text scene. According to the method, a CNN and an LSTM are combined, the advantages of the CNN and the LSTM are achieved at the same time, the problem that the CNN is weak in sequence correlation processing is solved, and the defect that the LSTM is insufficient in image feature extraction is overcome. According to the present invention, a new loss function calculation method CTC is combined to solvethe problem that the sample data is difficult to align during the text recognition process in a manner without alignment.

Description

technical field [0001] The invention belongs to the technical fields of computer vision, image processing and convolutional neural network, and in particular relates to an optical character recognition method in a patent text scene. Background technique [0002] With the continuous updating of computer hardware and software, and the gradual maturity of artificial intelligence (AI), it is of great practical significance to apply deep learning to the field of optical character recognition. Optical character recognition is to convert the text of various bills, newspapers, books, manuscripts and other printed materials into image information through optical input methods such as scanning, and then use text recognition technology to convert the image information into information that can be recognized by computers. Because there are too many influencing factors, including the habit of the writer, the printing quality of the document, the scanning quality of the scanner, the recog...

AI Technical Summary

Problems solved by technology

The main disadvantages are: (1) It takes a lot of time to do feature extraction. Usually, artificially designed features (such as histograms of directional gradients, etc.) are used to train character recognition models. Such single features can generalize quickly when fonts change. decline

(2) Excessive reliance on the result of character segmentation, in the case of overlapping and noise interference, the accuracy drops seriously

(3) Generally, good results can only be obtained in simple scenes, and poor results in complex scenes

LSTM can handle feature extraction of existing time series, however, traditional LSTM can only deal with short-term memory, because too long sequence will cause the gradient to disappear

Images