Efficient GANs for Document Image Binarization Based on DWT and Normalization

2024-07-05

arXiv.org

AI Summary

To address the high computational cost of training and inference in GAN-based document image binarization—particularly the trade-off between efficiency and text readability—this paper proposes a lightweight three-stage GAN framework. The method innovatively integrates discrete wavelet transform (DWT) with normalization mechanisms to design compact generator and discriminator architectures, and incorporates multi-scale adversarial training alongside a customized perceptual loss function. Experimental results demonstrate that the proposed approach maintains strong shadow suppression and noise robustness while significantly improving computational efficiency: training time is reduced by 10%, inference speed increases by 26%, and the Avg-Score reaches 73.79—matching state-of-the-art (SOTA) performance. This framework provides an efficient and practical solution for real-time document image enhancement.

Technology Category

• Computer Vision: Generative Adversarial Networks (GANs) for Vision
• Natural Language Processing: Generation
• Machine Learning: Deep Generative Models & Autoencoders

Application Category

• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines

Abstract

For document image binarization task, generative adversarial networks (GANs) can generate images where shadows and noise are effectively removed, which allow for text information extraction. The current state-of-the-art (SOTA) method proposes a three-stage network architecture that utilizes six GANs. Despite its excellent model performance, the SOTA network architecture requires long training and inference times. To overcome this problem, this work introduces an efficient GAN method based on the three-stage network architecture that incorporates the Discrete Wavelet Transformation and normalization to reduce the input image size, which in turns, decrease both training and inference times. In addition, this work presents novel generators, discriminators, and loss functions to improve the model's performance. Experimental results show that the proposed method reduces the training time by 10% and the inference time by 26% when compared to the SOTA method while maintaining the model performance at 73.79 of Avg-Score. Our implementation code is available on GitHub at https://github.com/RuiyangJu/Efficient_Document_Image_Binarization.

Problem

• Enhances color document images for better text recognition
• Reduces training and inference times of GAN-based binarization
• Uses multi-scale feature extraction with Haar wavelets for efficiency

Innovation

• Multi-scale feature extraction with Haar wavelet transformation
• Normalization before GAN training to reduce processing time
• Single GAN model for color document enhancement and binarization