Complex image processing with less data—Document image binarization by integrating multiple pre-trained U-Net modules. (January 2021)

• Record Type: Journal Article
• Authors: Kang, Seokjun; Iwana, Brian Kenji; Uchida, Seiichi
• Is Part Of: Pattern recognition. Volume 109(2021)
• Publication Date: 2021-01
• Subjects: Convolutional neural network -- U-Net -- Document image binarization -- DIBCO -- H-DIBCO; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2020.107577
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: Propose a novel document binarization method by cascading pre-trained U-Nets. Use pre-trained U-Net for solving a training image shortage problem. Study for optimal inter-module skip-connections between U-Net modules. Analyze the results of DIBCO images included various types of noise. Compare all DIBCO dataset (2009–2018) and show robust performance. Graphical abstract: Abstract: Artificial neural networks have been shown significant performance in various image-to-image conversion tasks. However, complex conversions often require a large number of images for model training. Therefore, we propose a convolutional model for image-to-image conversions using a pipeline of simpler image processing modules. To verify our proposed approach, we use a document image binarization as the task. Document image binarization is an important process that affects the accuracy of document analysis and recognition. In this paper, we propose a novel document binarization method called Cascading Modular U-Nets (CMU-Nets). CMU-Nets consist of pre-trained modular modules useful for overcoming the problem of a shortage of training images. We also propose a novel cascading scheme for improving overall cascading model performance. We verify the proposed model on all available Document Image Binarization Competition (DIBCO) and the Handwritten-DIBCO (H-DIBCO) datasets.