Request Correction

A Multi-Modal Recognition Framework for Ancient Books Integrating DoRA-DPO Text Recognition and YOLO Layout Analysis

The digitization and intelligent analysis of ancient Chinese documents face significant challenges due to diverse scripts, complex layouts, and the prevalence of rare characters. We present a comprehensive multi-modal recognition framework developed for the closed-modality track of the EvaHan 2026 Ancient Chinese Document Multi-Modal Recognition Shared Task. Our approach integrates two specialized pipelines to address these complexities. For text recognition (Tasks A and C), we propose a high-precision OCR system based on the domain-adapted Xunzi_Qwen2_VL_7B_Instruct, leveraging DoRA within a two-stage progressive curriculum learning strategy. To further refine character accuracy, DPO is incorporated alongside a dual-adapter architecture for rare character error localization and correction. For layout detection (Task B), we implement DocLayout-YOLO, enhanced by domain-specific pre-training and Mosaic augmentation to achieve efficient NMS-free element detection. Furthermore, a multi-round robust inference strategy, featuring automatic retry mechanisms and multi-prompt brute-force search, is introduced to handle stubborn and degraded samples effectively. Experimental results demonstrate that our proposed framework achieves superior performance across all evaluation metrics, highlighting its robustness and effectiveness in the digital preservation of ancient Chinese heritage.