Style-Aware Hierarchical Framework for Museum Artwork Recognition

Abstract

Automatic museum artwork identification faces significant challenges arising from extremely limited training samples per title—typically a single exemplar image per class—compounded by museum-specific imaging conditions and high intra-class visual variability. This paper presents a style-aware hierarchical recognition framework that addresses these challenges through three coordinated components. First, a domain-specific augmentation pipeline generates ten synthetic variants per original image, simulating spot lighting, lens distortion, motion blur, and partial occlusion. Second, the proposed Style-Aware Feature Fusion (SAFF) module is mounted atop an InceptionResNetV2 backbone; it extracts a 256-dimensional global style embedding via average pooling, generates channel-wise attention weights through a sigmoid-activated dense layer, and reweights the multi-scale feature map in a residual manner to emphasize stylistically discriminative cues such as brushstroke texture, color layering, and compositional rhythm. Third, a two-stage hierarchical classifier first assigns each artwork to one of four departments and then applies a department-specific title-level classifier, reducing the effective class space from 261 to between 42 and 86 classes per model. Evaluated on the Metropolitan Museum of Art benchmark, the framework achieves 86% accuracy at the department level and 84% weighted-average Top-1 accuracy across 261 title classes. The end-to-end hierarchical pipeline yields a final Top-1 accuracy of 75%, substantially outperforming flat baseline models including InceptionResNetV2 (56%), InceptionV3 (45%), and VGG16 (37%). These results demonstrate that combining domain-aware augmentation, style-sensitive feature fusion, and hierarchical decomposition provides a robust and scalable foundation for assistive museum applications.