| Abstract [eng] |
This study addresses the problem of clothing classification in the unbalanced DeepFashion dataset. It reviews existing methods for clothing classification and presents an improved pipeline that combines minority classes expansion with generative adversarial networks (GANs) and a YOLO11 classifier. The improvement is highlighted by a comparison of balancing strategies and other model variants. For the experiments, the original dataset of 289 222 garment images assigned with 46 labels was cleaned. The smallest classes were merged or discarded, leaving 24 classes. Every image was cropped to its bounding box and resized to 256x256 px. Two balancing routes were examined: standard augmentations and synthetic data generation using StyleGAN2-ADA and the newer version – StyleGAN3. StyleGAN3 converged faster and produced fewer artifacts than StyleGAN2-ADA. The classification stage evaluated a linear YOLO11 model alongside hierarchical variants. On the raw, imbalanced data the linear baseline achieved 76.15% top-1, 92.59% top-3 and 96.60% top-5 accuracy. Expanding minority classes with StyleGAN3 generated images raised performance to 77.53% top-1, 93.48% top-3 and 97.40% top-5. This GAN-balanced method outperformed both the augmentation-only and hierarchical YOLO11 methods, and surpassed the previously reported results on DeepFashion (91.99% top-3, 96.44% top-5). Results indicate that class distribution highly affects accuracy. Expanding smaller classes data with high quality GAN images delivers the best results while keeping the network simple. The project design section details the data preparation and modelling choices, while the experiments section presents the experiment results. |
