An Efficient CRNN Model with the Multi-scale Feature Fusion for Text Recognition from Chinese Medical Reports

The critical stage in extracting textual data from medical reports for subsequent processing is text recognition. However, the digitization of Chinese medical reports poses a significant challenge to the presence of multilingual texts and the intricate structure of Chinese characters, resulting in reduced recognition accuracy. We create a large-scale Chinese medical text image dataset (CMtext) and propose a Convolutional Recurrent Neural Network (MFF-CRNN) model with Multi-scale Feature Fusion for improving text recognition accuracy on Chinese medical reports. The model is enriched with three modules, namely residual network ResBlock, multi-scale feature fusion, and combined loss function, to enhance the extraction of multi-scale features at different convolutional layers and fuse them with CRNN. Multi-scale feature fusion is designed to combine low- and high-dimensional features in convolutional neural networks to improve feature extraction and enhance the robustness of the model. The ablation experiments demonstrate that the inclusion of each module is both beneficial and essential for achieving model improvement. MFF-CRNN achieves an accuracy of 98.9% in text recognition on the CMtext dataset, which outperformed existing state-of-the-art approaches. Additionally, the model has been tested on various datasets. The experimental results reveal its strong capability for generalization.