Hybrid CNN-ViT for Offline Handwritten Chinese Text Recognition with MSGA and Convolutional Attention

To address the core challenges of local feature sparsity and global structure information decay in offline handwritten Chinese text recognition, this study proposes a hybrid CNN-ViT architecture integrating a Multi-Scale Gated Aggregator (MSGA) and a Bidirectional Decomposed Convolutional Attention module. The framework uses a lightweight CNN to extract low-level features, employs MSGA to perform gated aggregation of multi-scale features for local detail enhancement, and utilizes the convolutional attention module to model vertical/horizontal spatial dependencies, optimizing the feature representation input to ViT. Experimental results show that the architecture achieves 98.64% Correct Rate (CR) and 97.92% Accuracy Rate (AR) on the CASIA-HWDB2.0-2.2 dataset, and 98.70% CR and 98.26% AR on TAL_OCR_CHN. Compared with mainstream methods, it significantly improves recognition accuracy in complex handwriting scenarios, verifying the synergistic effectiveness of local enhancement and global modeling, and providing a new technical path for line-level offline handwritten Chinese text recognition.