Lightweight Dongba character recognition: A novel and simple baseline

Dongba ancient books are listed as “World Memory Heritage” by UNESCO, which reflects its important status and significant impact on the study of world languages. Due to the lack of artificial prior knowledge, identifying Dongba characters with machines has become an important task in the field of historical and cultural protection. Existing methods generally construct complex network structures with numerous parameters to continuously improve the model performance. However, the huge memory consumption makes these models unsuitable for deployment in practical application scenarios. To fill this gap, we propose a novel lightweight information guidance network (LIGNet) for Dongba character recognition, aiming to make a better trade-off between network capacity and recognition accuracy. First, we propose a novel Lightweight Inverted Residual (LIR) block and a novel Lightweight Hybrid Attention (LHAtt) module. They ensure that the network maintains a low capacity level. Second, since the reduction of network parameters leads to limited representational capacity, we further propose a novel Forward Information Guidance (FIG) unit. This independent unit forms an auxiliary branch of the network, which strengthens the information transfer process of character features within the whole network. Extensive experimental results show that our LIGNet can favorably outperform existing methods on public benchmark datasets. Our code and models will be made available at https://github.com/DrChen215/LIGNet.