Imbalanced Classification via Feature Dictionary-based Minority Oversampling

Image classification research is one of the fields continuously studied in computer vision domain, and several related studies have been actively conducted until recently. However, a limit exists regarding prediction performance real-world datasets due to data imbalance problem between classes. Data augmentation through artificial sample generation for minority classes methods used overcome this limitation. Among various oversampling methods, we propose feature dictionary-based generative model method. Feature dictionaries are built pretrained extractor, proposed synthesizes samples based on dictionary. Class-to-class balanced training can be by fine-tuning classifier as additional class. We experiment applying framework fashion dataset, which has an extreme class imbalance. The experimental results demonstrate that achieved highest top-1 public datasets. In addition, analyze number dictionary test effectiveness elements comprise using ablation studies.