Fine-Grained Fashion Classification with Limited Labeled Data Via Generative Augmentation

초록

Fine-grained classification of fashion items is essential for enhancing user experiences in online shopping, enabling more effective categorization and personalized recommendations. While recent advances in Artificial Intelligence (AI) have shown promise, achieving high classification accuracy typically requires large volumes of labeled data—an expensive and labor-intensive requirement, particularly burdensome for smaller retailers. Existing approaches attempt to mitigate this challenge by pretraining models on synthetic or geometric data and fine-tuning on limited fashion images, but these methods have thus far plateaued at around 90% accuracy, falling short of practical deployment standards. In this paper, we introduce a novel iterative image generation framework designed to overcome data scarcity and significantly improve classification accuracy. Our method combines a conditional Generative Adversarial Network (cGAN) with a ResNet50-based image classifier in a closed-loop system. The cGAN generates synthetic fashion images conditioned on class labels, while the classifier filters outputs based on confidence scores and predefined criteria. This cycle is repeated iteratively, progressively enriching the training dataset with high-quality synthetic images and refining the classifier. We validate our approach on a task involving classification of five distinct neckline types. The converged model achieves an average accuracy of 94.6%, substantially outperforming previous methods despite using a limited amount of real labeled data. These results highlight the effectiveness of iterative data augmentation using generative models for fine-grained visual classification in resource-constrained settings. © 2020 Tsinghua University Press.

키워드