Meta-learning-optimized deep feature fusion strategies for few-shot classification of rare Chinese figure painting styles

Chinese figure painting, with its rich history and diverse styles, represents a significant cultural heritage that captures the essence of Chinese art and aesthetics. However, many traditional figure painting styles, particularly rare ones, remain underrepresented and face considerable challenges in modern-day recognition, categorization, and preservation. This issue is especially prominent in the context of few-shot learning, where models must classify and identify rare categories with only minimal training data. To address this challenge, the research proposes a novel meta-learning-optimized deep feature fusion strategy for the few-shot classification of rare Chinese figure painting styles. The research utilizes a Chinese art styles dataset of Chinese figure paintings, with a specific focus on rare and fine-grained styles. Preprocessing steps such as image resizing and noise removal were employed to enhance data quality. The ResNet-50 model was utilized to extract low-level characteristics from the input images, such as edges, textures, and colors. The features extracted from different layers were then fused using deep feature fusion techniques to create a comprehensive and discriminative representation of the painting styles. This feature fusion process was further optimized through a meta-learning approach based on the Elephant Clan Optimized Attention-based Variational Autoencoder (ECO-Att-VAE) method, which fine-tunes feature weighting based on the support set in the few-shot classification task. Experiments show that the suggested strategy works noticeably better than current methods in the categorization of unusual art styles. Results show that the model achieves a notable improvement in classification accuracy, providing a robust and effective solution for few-shot classification in the domain of rare Chinese figure painting styles.