DocRect: A Transformer-Based Approach for Distorted Document Image Correction

Abstract Document images captured by smartphones often suffer from distortions due to physical deformations like wrinkles or folds and perspective issues from camera positioning. To address this, we propose DocRect, a novel framework for document image rectification. DocRect employs a phased approach, integrating a background mask generator, a primary feature extractor, a dewarping transformer, and a refinement module. Unlike standard vision transformers, it uses a symmetric encoder-decoder architecture to capture both local and global distortion features. The transformer decoder, paired with a learnable refinement module, generates a pixel-wise backward rectification flow field, enabling precise mapping from distorted to undistorted images. Quantitative evaluations on the DocUNet Benchmark dataset demonstrate that DocRect achieves superior performance in image similarity and OCR accuracy compared to existing methods.