An Explainable AI Classification Framework for Impressionist Paintings Based on Fourier Transform Feature Disentanglement: Toward Human–AI Collaborative Art Appreciation

Abstract

Impressionist artworks, marked by distinctive brushstrokes and light effects, pose challenges for holistic deep learning models. Existing classification methods often lack interpretability and mathematical rigor. To address this, we propose an explainable AI framework for Impressionist painting classification based on feature disentanglement. The framework uses Fourier transform to decompose images into magnitude and phase spectra and then applies frequency filtering to separate high-frequency components (brushstroke textures, edges) from mid-to-low frequencies (global composition, color transitions). A state space model captures long-range dependencies within frequency representations. Disentangled features are fused with the original image features via a dual-branch injection mechanism. Experiments show that a lightweight model with 10M parameters, using this injection strategy, outperforms conventional baselines with 25M parameters. Moreover, the approach provides interpretability by explicitly linking classification decisions to brushstroke organization, frequency energy distribution, and semantic composition. This work offers a unified, computationally efficient, and mathematically rigorous solution for feature disentanglement in artistic image analysis, laying a theoretical foundation for explainable computational aesthetics and human–AI collaborative appreciation.