Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ShaDocFormer: A Shadow-Attentive Threshold Detector With Cascaded Fusion Refiner for Document Shadow Removal

Weiwen Chen, Yingtie Lei, Shenghong Luo, Ziyang Zhou, Mingxian Li, Chi-Man Pun

TL;DR

ShaDoc-Former is proposed, a Transformer-based architecture that integrates traditional methodologies and deep learning techniques to tackle the problem of document shadow removal and outperforms current state-of-the-art methods in both qualitative and quantitative measurements.

Abstract

Document shadow is a common issue that arises when capturing documents using mobile devices, which significantly impacts readability. Current methods encounter various challenges, including inaccurate detection of shadow masks and estimation of illumination. In this paper, we propose ShaDocFormer, a Transformer-based architecture that integrates traditional methodologies and deep learning techniques to tackle the problem of document shadow removal. The ShaDocFormer architecture comprises two components: the Shadow-attentive Threshold Detector (STD) and the Cascaded Fusion Refiner (CFR). The STD module employs a traditional thresholding technique and leverages the attention mechanism of the Transformer to gather global information, thereby enabling precise detection of shadow masks. The cascaded and aggregative structure of the CFR module facilitates a coarse-to-fine restoration process for the entire image. As a result, ShaDocFormer excels in accurately detecting and capturing variations in both shadow and illumination, thereby enabling effective removal of shadows. Extensive experiments demonstrate that ShaDocFormer outperforms current state-of-the-art methods in both qualitative and quantitative measurements.

ShaDocFormer: A Shadow-Attentive Threshold Detector With Cascaded Fusion Refiner for Document Shadow Removal

TL;DR

ShaDoc-Former is proposed, a Transformer-based architecture that integrates traditional methodologies and deep learning techniques to tackle the problem of document shadow removal and outperforms current state-of-the-art methods in both qualitative and quantitative measurements.

Abstract

Document shadow is a common issue that arises when capturing documents using mobile devices, which significantly impacts readability. Current methods encounter various challenges, including inaccurate detection of shadow masks and estimation of illumination. In this paper, we propose ShaDocFormer, a Transformer-based architecture that integrates traditional methodologies and deep learning techniques to tackle the problem of document shadow removal. The ShaDocFormer architecture comprises two components: the Shadow-attentive Threshold Detector (STD) and the Cascaded Fusion Refiner (CFR). The STD module employs a traditional thresholding technique and leverages the attention mechanism of the Transformer to gather global information, thereby enabling precise detection of shadow masks. The cascaded and aggregative structure of the CFR module facilitates a coarse-to-fine restoration process for the entire image. As a result, ShaDocFormer excels in accurately detecting and capturing variations in both shadow and illumination, thereby enabling effective removal of shadows. Extensive experiments demonstrate that ShaDocFormer outperforms current state-of-the-art methods in both qualitative and quantitative measurements.
Paper Structure (15 sections, 5 equations, 5 figures, 2 tables)

This paper contains 15 sections, 5 equations, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Traditional Methods
  4. Learning-based Methods
  5. Methodology
  6. Overview
  7. Shadow-attentive Threshold Detector
  8. Cascaded Fusion Refiner
  9. Objective Function
  10. Experiments
  11. Datasets and Metrics
  12. Experiment Settings
  13. Comparisons with State-of-the-Arts
  14. Ablation Studies
  15. Conclusion

Figures (5)

  • Figure 1: The visual results of the input document shadow image (a), along with the results of the traditional method (b), learning-based methods (c) and (d), ours (e), and the target (f). Our model effectively removes the shadow while preserving the content and color information of the original document.
  • Figure 2: Model diagram of the proposed Cascaded Fusion Refiner framework and the Convolutional Depth Grouped Fusion Net framework.
  • Figure 3: Model diagram of the proposed Shadow-attentive Threshold Detector framework. It combines the intricacies of convolutional layers with the contextual prowess of Transformer blocks for enhanced image processing.
  • Figure 4: The qualitative results of comparing the methods on the RDD and Kligler dataset.
  • Figure 5: The visual results of the input document shadow image (a), along with the results of our model without EBO (b), without STD (c) , without CDGF (d), our full model (e) and the target (f). All the elements within our model play a crucial role in its success, and we found that neglecting even a single component led to a marked decline in the performance indicators we tracked. This outcome highlights the necessity of employing a holistic strategy to attain the superior outcomes our method has shown.