SSD-DIS: A Semi-Synthetic Shadow Dataset for Document Images

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Smartphones are increasingly used for digitizing physical documents, driven by remote work and education needs.
• Shadows in captured document images significantly impair readability and information extraction.
• Existing shadow removal datasets have limitations for training robust AI models.

Purpose of the Study:

• To develop a novel semi-synthetic dataset (SSD-DIS) for improving document shadow removal.
• To address the limitations of current datasets in simulating realistic shadow conditions.
• To enhance the performance of AI models in document shadow removal tasks.

Main Methods:

• A semi-synthetic dataset (SSD-DIS) comprising 12,224 image sets was created.
• Blender software was utilized to generate shadow masks and incorporate multi-source shadow-free images.
• Shadow intensity and color were adjusted to simulate diverse real-world scenarios.

Main Results:

• Experiments demonstrated that SSD-DIS effectively enhances neural networks' ability to learn document shadow features.
• Models trained on SSD-DIS exhibited superior performance compared to those trained on traditional datasets.
• The dataset supports advancements in research for document shadow removal algorithms.

Conclusions:

• The proposed SSD-DIS dataset is a valuable resource for advancing document shadow removal technology.
• Training AI models on SSD-DIS leads to improved performance in handling shadows in document images.
• This work contributes to better usability of digitized documents in remote contexts.