A Chinese scene text line identification method based on residual convolution and a recurrent neural network

Abstract

The invention discloses a Chinese scene text line identification method based on residual convolution and a recurrent neural network. The method comprises the following steps: collecting a Chinese scene text training image, performing normalization processing on the size of the training image, performing data augmentation processing on the training image, designing a residual convolutional neuralnetwork, a residual recurrent neural network and a CTC model, training horizontal text lines and vertical text lines, and selecting a result with higher confidence as an identification result. According to the invention, the convolutional neural network and the recurrent neural network are combined; the problem of Chinese scene text line identification is solved; error recognition caused by character segmentation and error segmentation on the text lines is avoided; the training of the residual error connection acceleration model is added into the convolutional neural network and the recurrentneural network, so that a practical Chinese scene text recognition model is obtained, the robustness is high, and Chinese text lines with complex backgrounds, complex illumination and various fonts can be recognized.

Description

technical field [0001] The present invention relates to the field of computer vision, in particular to a Chinese scene text line recognition method based on residual convolution and recursive neural network Background technique [0002] Text is a basic tool for human beings to communicate and understand information. With the popularization of smart phones and the rapid development of the Internet, it has gradually become a very popular way of life to obtain, process and share information through mobile terminal cameras such as mobile phones and tablet computers. , generally speaking, in a scene where text and other objects coexist, users tend to pay more attention to the text content in the image, and the text content is also very important for the understanding of the image, so how to accurately and quickly identify the text in the image Text, which will have a deeper understanding of the user's intention of shooting and the theme of the work. [0003] Compared with tradit...

AI Technical Summary

Problems solved by technology

[0003] Compared with traditional document images, the text background in natural scenes is more complex, the fonts are diverse, and it is easily affected by lighting and shooting angles, so the recognition is very difficult. In the early research on scene text recognition, it is usually necessary to convert the text Segment the words in the line, then extract the features of the word image, and finally identify them through a word classifier. In this type of method, the accuracy of recognition is highly dependent on the accuracy of the previous word segmentation. If the word is incorrectly segmented, then It is difficult to identify correctly. In addition, the artificially designed features cannot well represent the features of each level of the image. In recent years, deep ...

Images