Item-region-based style classification network (IRSN): a fashion style classifier based on domain knowledge of fashion experts.

Fashion style is expressed through the way clothing and accessories are put together, as well as the silhouettes, textiles, colors, and shape details of each fashion item. The challenge of style classification lies in the wide visual variation within the same style and the existence of visually similar styles. Fashion experts categorize fashion styles not only by global appearance but also by the attributes of individual items and their combinations. We propose an item-region-based fashion style classification network (IRSN) that effectively classifies fashion styles by analyzing item-level features and their combinations. IRSN extracts item features using item region pooling (IRP), analyzes them separately, and aggregates them using gated feature fusion (GFF). In addition, IRSN applies a dual-backbone architecture that combines a domain-specific feature extractor and a general feature extractor pretrained with a large general image-text dataset. In the experiment, we evaluated IRSN variants based on six widely used backbones, including EfficientNet, ConvNeXt, and SwinTransformer. The IRSN models outperformed their baseline models by an average of 8.9% and a maximum of 16.7% on the FashionStyle14 dataset, and by an average of 9.4% and a maximum of 17.0% on the ShowniqV3 dataset. The visualization results support that the IRSN models are more effective than the baseline models in capturing differences between similar style classes. [ABSTRACT FROM AUTHOR]