Multimodal Framework for Cursive Calligraphy Recognition with Unet, ResNet, and LLMs

Cursive calligraphy recognition poses significant challenges due to its inherent complexity and stylistic diversity, where traditional machine learning methods often fail to achieve efficient recognition. This paper proposes a robust recognition framework integrating Unet semantic segmentation, ResNet feature extraction, and Large Language Models (LLMs). The framework first employs Unet for precise Chinese character segmentation, extracts discriminative features through ResNet variants (ResNet18/50/101), and leverages LLM-based contextual semantic correction. Experimental results demonstrate that the proposed algorithm improves accuracy on multi-style cursive datasets while optimizing computational efficiency through model quantization and multimodal fusion. This work provides technical references for ancient text restoration and document digitization.