DocQT: Improving Document Forgery Localization Robustness via Diverse JPEG Quantization Tables
Researchers have developed DocQT, a new dataset and method to improve the robustness of document forgery localization models. These models often fail in real-world scenarios due to a mismatch between training data and operational document compression. DocQT addresses this by using diverse JPEG quantization tables sampled from real-world insurance documents, leading to significant gains in localization accuracy and reduced false positives, particularly for architectures that explicitly process quantization table information. AI
IMPACT Enhances the reliability of AI models used for detecting manipulated documents in real-world applications.