Dynamically Adaptive Deformable Feature Fusion for multi-scale character detection in ancient documents

초록

Robust character detection of various sizes in ancient East-Asian handwritten documents is essential for accurate classification and translation. While prior studies have addressed various challenges associated with the writing styles and page layouts of historical documents, they struggle to detect small-sized characters, especially with an area below 32 square units due to their limited presence in the data corpus. Furthermore, the physical degradation, presence of artifacts, and inconsistencies in text density complicate character detection of historical documents across multiple scales. In this study, we propose a novel multi-scale character detection named Dynamically Adaptive Deformable Feature Fusion (DAF). This approach leverages deformable convolutions to improve feature extraction for the complex and irregular shapes found in ancient East-Asian manuscripts. We also present an innovative Adaptive Weight Module that dynamically adjusts top-down features across different levels by utilizing trainable weights. This enhances the detection of multi-scale characters and effectively identifies small-sized characters within documents. Further, we contribute to existing research by proposing a set of detection metrics, specifically designed to evaluate both general and scale- specific detection scenarios. Extensive experiments conducted on several datasets of ancient handwritten documents including the Nancho dataset, Multiple Tripitaka in Han dataset (MTHv2), and Kuzushiji dataset demonstrate the superior performance of our proposed DAF framework over existing multi-scale detection methods.

키워드