Abstract:Buddha statues, as a symbol of many religions, have significant cultural implications that are crucial for understanding the culture and history of different regions, and the recognition of Buddha statues is therefore the pivotal link in the field of Buddha study. However, the Buddha statue recognition requires extensive time and effort from knowledgeable professionals, making it a costly task to perform. Convolution neural networks (CNNs) are inherently efficient at processing visual information, but CNNs alone are likely to make inaccurate classification decisions when subjected to the class imbalance problem. Therefore, this paper proposes an end-to-end automatic Buddha statue recognition model based on saliency map sampling. The proposed Grid-Wise Local Self-Attention Module (GLSA) provides extra salient features which can serve to enrich the dataset and allow CNNs to observe in a much more comprehensive way. Eventually, our model is evaluated on a Buddha dataset collected with the aid of Buddha experts and outperforms state-of-the-art networks in terms of Top-1 accuracy by 4.63\% on average, while only marginally increasing MUL-ADD.