ArtNAS-Style: Style-Prior-Guided Differentiable Architecture Search for Artwork Recognition

Abstract

Recognizing visual artworks is difficult because images of art combine large within-class diversity, ambiguous boundaries, and non-photographic textures with intertwined cues related to the artist, genre, style, and historical period. Standard convolutional or transformer models are commonly inherited from natural-image classification, and such backbones may fail to emphasize the visual statistics that distinguish paintings, drawings, prints, and other artistic media. This paper presents ArtNAS-Style, a differentiable neural architecture search framework that embeds artistic style priors into the architecture search stage for visual art recognition. Rather than optimizing architecture choices only through classification loss, the framework employs a style-prior-aware bilevel objective that jointly accounts for recognition accuracy, diversity of style-sensitive features, preservation of brushstroke-like texture, and robustness across stylistic domains. The proposed Style Prior Encoding Module derives Gram-statistic, spectral, and palette-distribution cues from each artwork, and the Prior-Conditioned Search Cell uses these cues to adaptively modulate candidate operations during the search. A Style-Balanced Architecture Regularizer further discourages the discovered architecture from specializing to dominant artistic categories and promotes discriminative features across heterogeneous art domains. The search model is compact, trainable end to end, and transferable to multiple recognition tasks. Our evaluation protocol covers style, artist, school, and object-type recognition, in addition to parameter-matched ablations, robustness, statistical testing, and computational costs.