TL;DR
This paper introduces a novel deep learning approach for removing shadows from document images, a common issue that degrades readability. By combining a Sobel edge prior with ground truth mask supervision, their method achieves efficient and effective shadow removal, outperforming existing techniques.
What Problem Does This Paper Address?
Digitizing documents often results in shadows that significantly impair the clarity and readability of the digital copies. While various methods exist for shadow removal, traditional techniques can be sensitive to noise and require extensive manual tuning, making them impractical for diverse scenarios. Existing deep learning approaches, though promising, often suffer from a large information burden, leading to high computational costs. This paper aims to bridge this gap by proposing an efficient deep learning model that can achieve superior shadow removal effects with reduced computational overhead.
Methods
The proposed architecture consists of two main components: MaskNet for mask prediction and FusionNet for feature fusion.
- MaskNet: This network takes a 4-channel feature map as input and predicts a pixel-wise local mask. It utilizes a gate block to divide the input into latent channels and then processes a 1-channel latent output through a sigmoid activation function to generate a mask ranging from 0 to 1. This mask prediction is guided by ground truth masks. The predicted mask is then combined with the latent features to produce residual images, which are fused across scales to form the final predicted mask.
- FusionNet: This component facilitates feature fusion. It takes the original feature map ($F$) and the feature map with the mask ($F - F_m$) as inputs. Kernel Generation Units (KGUs) synthesize weight tensors of different sizes (3x3, 7x7, 11x11) inspired by densely connected architectures. These KGUs enhance feature propagation and representation, leading to improved feature fusion. Kernel Transformation Units (KTUs) are also employed to generate convolutional kernels with varying dilation levels.
A key novelty is the incorporation of a Sobel text prior. This prior knowledge captures regular information about document structure and shadow formation, specifically leveraging edge information to guide the shadow removal process. This, combined with mask supervision, helps the model learn more quickly and efficiently.
Datasets
The authors evaluate their method on Jung's dataset and Kligler's dataset. These datasets are used for both training and evaluation.
Key Insights and Ideas
- Sobel Prior for Efficiency: Integrating a Sobel text prior is crucial for enhancing efficiency. By leveraging edge information inherent in documents, the model can learn more effectively and reduce the amount of data it needs to process.
- Mask Supervision for Precision: Using ground truth masks for supervision directly guides the MaskNet in predicting accurate masks, which are essential for precise shadow removal.
- Feature Fusion for Robustness: The FusionNet's approach to combining features from the original map and the masked map, using KGUs and KTUs, allows for a more comprehensive understanding of the image content, leading to better shadow removal.
Key Results
The paper evaluates its performance using standard metrics: Root Mean Square Error (RMSE), Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity (SSIM). While specific quantitative results are not detailed in the provided text, the authors claim their proposed method achieves "remarkable performance" and is more efficient than existing deep learning approaches. They compare their method against a comprehensive list of traditional and state-of-the-art deep learning baselines, including Kligler, AEFNet, BEDSR-Net, DC-ShadowNet, DHAN, LG-ShadowNet, Mask-ShadowGAN, SG-ShadowNet, ST-CGAN, and ShadowFormer.
Limitations and Future Work
The authors acknowledge that further work is needed to optimize their model for better memory usage and performance. They also plan to incorporate more datasets, such as the SD7K and ISTD datasets, to prove their model's robustness. Future research directions include developing more efficient and effective methods for document image processing, contributing to advancements in computer vision and document analysis.
Summary updated