Picture correction method and device based on neural network, equipment and medium

Abstract

The invention relates to artificial intelligence, and provides a picture correction method and device based on a neural network, equipment and a medium, and the method comprises the steps: cutting a to-be-recognized picture according to a cutting strategy, obtaining a cut sub-picture set, carrying out the convolution of each cut sub-picture through a convolution kernel, obtaining a feature graph set, and forming a one-dimensional sequence corresponding to the cut sub-pictures; obtaining a one-dimensional sequence corresponding to each segmented sub-picture, then obtaining a position coding vector of each segmented sub-picture in a segmented sub-picture set, connecting the position coding vector of each segmented sub-picture with the corresponding one-dimensional sequence to obtain an updated one-dimensional sequence corresponding to each segmented sub-picture, and finally, inputting each updated one-dimensional sequence into a Transform model for coding and decoding, and obtaining a decoding result. According to the method, after the to-be-recognized picture is converted into the one-dimensional sequence through cutting and convolution, more global information in the one-dimensional sequence is extracted by using the Transform model, so that the global semantic information of the picture resolution loss caused by down-sampling is avoided, and a better detail correction effect is achieved.

Description

technical field [0001] The invention relates to the field of image recognition of artificial intelligence, in particular to a neural network-based image correction method, device, computer equipment and storage medium. Background technique [0002] With the development of technology and the improvement of people's living standards, using mobile devices to take pictures to record document information is a common method today. However, mobile device shooting is usually affected by factors such as angle tilt, physical distortion of documents, deformation, etc., which brings great challenges to text recognition and structured information archiving. Automatically flattening a distorted document image can not only improve the accuracy of text recognition, but also greatly reduce the difficulty of extracting structured information, thus greatly improving the accuracy of document structured filing as a whole. [0003] For the correction method of distorted and skewed documents, the...

AI Technical Summary

Problems solved by technology

[0004] The embodiment of the present invention provides a neural network-based picture correction method, device, computer equipment, and storage medium, aiming to solve the correction method for distorted and skewed documents in the prior art. The current popular solution in the industry is based on 2D correction The method, the use of downsampling will damage the image resolution, resulting in the loss of local details, thus affecting the image correction effect

Images