Resolving Stylistic Overlap in Art Movements Using Semantic Embedding and Uniform Manifold Approximation and Projection

The growing availability of large-scale digitized fine art collections has opened new avenues for research in computational art analysis. Among these, accurately identifying the art movement of a painting is crucial for tasks such as indexing vast art databases. However, this task is complicated by the significant stylistic overlap between various art movements. This generated a lower F1 score in multiclass classification. According to our observation, there is a lack of High-Level Features (HLF) that capture the semantic information of the paintings. In this work we have proposed semantic analysis as a High level features. To obtain HLF, we employed a novel methodology based on a pipeline utilizing the Visual Language Model (VLM), Sentence Transformer, and Uniform Manifold Approximation and Projection for Dimension Reduction (UMAP). While classifying art movements (Abstract Expressionism, Color_Field_Painting, Early_Renaissance, Expressionism, Impressionism, Minimalism, and Realism) from the WikiArt dataset, we achieved a 98% recognition rate in Impressionism, which is a significant improvement over state-of-the-art classification models and achieved 75% overall accuracy, in the recognition of art movements.