A Multi-Stage System for Ancient Chinese OCR and Layout Understanding in the EvaHan2026 Shared Task

Abstract

This paper presents a multi-stage system for the EvaHan2026 shared task, addressing the complex challenges of ancient Chinese optical character recognition (OCR) and layout understanding. For text recognition (Tasks A and C), we adopt parameter-efficient LoRA fine-tuning on the Qwen2.5-VL-7B-Instruct vision-language model (VLM). By directly processing full-resolution long-column images, we preserve critical spatial and contextual integrity without heuristic region cropping. For document layout analysis (Task B), we propose a novel hybrid perception-reasoning paradigm. Instead of relying solely on scaling visual detectors, we decouple localization and understanding: utilizing a YOLO-based ensemble for precise spatial bounding, and casting the VLM as a semantic verifier to eliminate spurious detections. Evaluated on the official unseen test set, our system achieves substantial improvements over the provided baselines, obtaining a 0.0441 Character Error Rate (CER) for printed OCR, a 0.0793 CER for handwritten OCR (including variants), and a 0.5118 mAP@[0.5:0.95] for layout detection. These results demonstrate that integrating VLM-based semantic reasoning into traditional visual detection pipelines is highly effective for multimodal historical document analysis.

Resources

Details