| Title |
Improving DeepFashion dataset classification accuracy with StyleGAN2-ADA: addressing imbalanced data in fashion image recognition |
| Authors |
Moleikaitytė, Agnė ; Paulauskas, Jonas |
| Full Text |
|
| Is Part of |
CEUR Workshop proceedings: IVUS 2025: proceedings of the 30th international conference on information society and university studies (IVUS 2025), Kaunas, Lithuania, 15 May 2025 / edited by: I. Veitaitė, A. Lopata, T. Krilavičius, M. Woźniak.. Aachen : CEUR-WS. 2026, vol. 4213, p. 191-200.. ISSN 1613-0073 |
| Keywords [eng] |
DeepFashion ; StyleGAN2-ADA ; Data augmentation ; Imbalanced dataset ; Fashion image recognition |
| Abstract [eng] |
Fashion classification remains challenging due to the high variability in apparel styles and the significant class imbalance in fashion datasets. This study utilizes the DeepFashion dataset, which consists of 289,222 images across 46 clothing categories, to evaluate the impact of data preprocessing, augmentation, and synthetic data generation on classification performance. To address class imbalance, underrepresented categories were merged with similar classes or supplemented with additional training data. Two dataset expansion techniques were explored: traditional augmentation (flipping, rotation, color jittering, and Gaussian blur) and synthetic image generation using StyleGAN2-ADA. The YOLO11 model was employed for classification, and its performance was assessed using top-k accuracy metrics. Experimental results show that preprocessing alone improved classification accuracy, while adding synthetic data further enhanced model performance, achieving a top-3 accuracy of 93.44% and a top-5 accuracy of 97.34%. Analysis of the confusion matrix revealed that while synthetic data helped mitigate class imbalance, misclassifications persisted among visually similar categories. These findings highlight the potential of generative models in enriching training datasets and improving classification performance. However, further feature extraction and inter-class discrimination refinements are necessary for optimal results. |
| Published |
Aachen : CEUR-WS |
| Type |
Conference paper |
| Language |
English |
| Publication date |
2026 |
| CC license |
|
