Scholarlink

<p>The deblurring of blurred-text images is a critical task in computer vision, especially in applications such as document digitization, OCR (Optical Character Recognition), and image restoration. Traditional deblurring methods often struggle with text images due to the complex structures and high-frequency details inherent in textual content. However, existing datasets are often insufficient to train robust deblurring models due to their limited variety in blur patterns, text formats, and environmental conditions. This insufficiency poses significant challenges in data augmentation, where traditional techniques may distort text structures or fail to simulate the complexity of real-world blurs, leading to models that struggle to generalize effectively. To address these issues, we propose a novel data augmentation method using existing dataset SROIE along with Non-Euclidean extrapolation, designed to create more realistic and varied training examples. It consists the combinations of existing Riemannian mixup (R-Mixup), Hyperbolic Spaces for generating a new dataset <Emphasis Type="BoldItalic">SROIE-RH</Emphasis> and the combination of Hyperbolic Spaces, and Gaussian mix-up for creation another dataset <Emphasis Type="BoldItalic">SROIE-HG</Emphasis> to improve the quality and the sources of the augmented dataset using. The validity of these methods is proven through various tests and shows enhanced increases in the model’s resilience and precision leading to sharper and more accurate deblurring of blurred-text images.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Data augmentation for blurred-text image deblurring using non-euclidean extrapolations

  • 期刊论文

  • 发表日期：

    2025年1月21日

查看全文
  • Arti Ranjan,
  • M. Ravinder

摘要

The deblurring of blurred-text images is a critical task in computer vision, especially in applications such as document digitization, OCR (Optical Character Recognition), and image restoration. Traditional deblurring methods often struggle with text images due to the complex structures and high-frequency details inherent in textual content. However, existing datasets are often insufficient to train robust deblurring models due to their limited variety in blur patterns, text formats, and environmental conditions. This insufficiency poses significant challenges in data augmentation, where traditional techniques may distort text structures or fail to simulate the complexity of real-world blurs, leading to models that struggle to generalize effectively. To address these issues, we propose a novel data augmentation method using existing dataset SROIE along with Non-Euclidean extrapolation, designed to create more realistic and varied training examples. It consists the combinations of existing Riemannian mixup (R-Mixup), Hyperbolic Spaces for generating a new dataset SROIE-RH and the combination of Hyperbolic Spaces, and Gaussian mix-up for creation another dataset SROIE-HG to improve the quality and the sources of the augmented dataset using. The validity of these methods is proven through various tests and shows enhanced increases in the model’s resilience and precision leading to sharper and more accurate deblurring of blurred-text images.