Pushing the Boundaries of Chinese Painting Classification on Limited Datasets: Introducing a Novel Transformer Architecture with Enhanced Feature Extraction.

Haiming Zhao,Jiejie Chen,Ping Jiang,Tianrui Wu,Zhuzhu Zhang

DOI: https://doi.org/10.1007/978-981-99-8138-0_15

2023-01-01

Abstract:This study ventures into the relatively unexplored field of Chinese painting classification. The paper addresses challenges such as limited datasets, feature map redundancy, and inherent limitations of the Transformer model by curating a high-resolution dataset of Chinese paintings from various dynasties. To enhance efficiency, we introduce a strategy for selecting channels that capture unique features of these paintings and propose a novel architecture. This architecture uses local self-attention with sliding windows and two residual connections for spatial token mixing and channel feature transformation. Experiments show our architecture’s superior accuracy in classifying small datasets. The research advances the field of Chinese painting classification, demonstrating deep learning models’ potential in this area. The dataset and code can be accessed at: https://github.com/qwerty0814/gangan .