Training method for correction model of distorted document picture and correction method of distorted document picture

Abstract

The invention discloses a training method for a correction model of a distorted document picture and a correction method for the distorted document picture. The method comprises the following steps: obtaining a plurality of flat document pictures, and carrying out deformation of the flat document pictures to obtain training document pictures; and taking the training document pictures and the flatdocument pictures corresponding to the training document pictures as a training data set of deep learning to perform model training to obtain a correction model, and performing end-to-end correction on the distorted documents by utilizing the correction model to obtain a relatively flat document picture. According to the invention, the training picture is obtained by carrying out random deformation on the flat document picture; under the condition, whether the model obtained through training has good adaptability or not can be judged through picture comparison or other automatic modes, so thatthe model training efficiency and the model selection efficiency can be greatly improved, the high quality of training materials in the document processing process is improved, and a basis is provided for obtaining a high-quality correction model.

Description

technical field [0001] The present application relates to a training method for a correction model of distorted document pictures and a correction method for distorted document pictures. Background technique [0002] Due to the popularity of mobile phones, taking photos at hand has become a common method for digitally recording documents, and subsequent operations such as text recognition can be performed based on this. However, due to the often distorted and / or distorted documents obtained from random photos, the lighting conditions are also unstable. Character recognition under this situation is difficult to achieve the desired effect. In order to improve the recognition efficiency, flatbed scanners are generally used to ensure the regularity of documents. But it is conceivable that this kind of equipment is not easy to carry, and the cost is also high. Therefore, how to solve distorted document images to facilitate content analysis and information extraction becomes qu...

AI Technical Summary

Problems solved by technology

The first thing to deal with is to restore the document, but the existing restoration method is mostly a kind of stretching and twisting of the image, which cannot make up for the defect in the document picture, and sometimes expands the defect, making some text more difficult to recognize Or the recognition error rate increases sharply

And a common practical problem when taking images of documents is that the documents are not in ideal condition for scanning: they may be warped, folded or shrunk

Images