Distortion Document Rectification Based on Offset Field Learning
DONG Qianqian
CHEN Liang
WANG Xinxin
Abstract：The rectification of distorted document images is an essential approach to enhance text recognition and information extraction accuracy,thereby promoting the intelligentization of documents.Existing methods for correcting distorted documents pri-marily focus on tightly cropped document images and are ineffective in addressing documents with significant environmental bound-aries.To address this issue,this paper proposes a distortion document correction method that involves detection before correction.Initially,improvements to the Mask R-CNN are introduced to enhance the accuracy of detecting blurry document images.This in-cludes the incorporation of a higher-resolution ROI Align layer and decoder layer,along with optimizations to the Mask module.Subsequently,precise correction of distorted documents is achieved by introducing offset field learning.Experimental results demon-strate notable performance on the Doc3DShade dataset compared to other state-of-the-art algorithms,with a 1.6％improvement in SSIM metric,a 0.5％decrease in LD metric,and a 5.42％reduction in character error rate.These findings substantiate its excep-tional performance and provide an efficient and accurate document correction solution for document capture on mobile devices.
Keywords：offset fieldimage rectificationMask R-CNNmobile devices
Publication Date：2025-12-20
Online Publishing Date：2026-03-23(First online date of this platform, not the publication date of the document)
Pages：5( 41-45 )