Art Painting Classification Based on an Improved Dual-Branch Vision Transformer Model
With the advancement of art digitization,there is an urgent need for intelligent systems capable of accurately analyzing and organizing art painting collections and uncovering relationships between different artistic attributes based on visual elements of the paintings.To this end,a deep learning method is proposed for improving art painting classification using an improved Vision Transformer model and feature optimization algorithm.First,features are extracted from art painting images using an improved dual-branch Vision Transformer(CrossViT).Shared features are extracted through a dual-branch architecture,achieving multi-scale feature representations.A cross-task fusion phase is designed,where specific task tokens are processed with separate branches,and information is exchanged through a cross-attention module.Subsequently,a Chaos Game Optimization(CGO)algorithm combined with the Nutcracker Optimizer(NO)is employed to determine an optimal subset of features.The effectiveness of the proposed improved CGO algorithm is vali-dated through algorithmic tests on eight functions from the CEC2022 benchmark.Additionally,experimental results on the SemArt dataset for tasks including type,genre,and period recognition in art painting classification demonstrate that the proposed method accurately accomplishes art painting classification based on various task requirements,outperfor-ming other state-of-the-art methods.
Art painting classificationDeep learningVision TransformerChaos Game OptimizationNutcracker Opti-mizer
NETL
NSTL
万方数据
