The rapid expansion of online fashion retail has led to a massive increase in product images, making accurate clothing classification a crucial task. Misclassification not only disrupts customer satisfaction but also contributes to higher return rates and operational inefficiencies. Effective apparel classification enhances search results, improves personalized recommendations, and optimizes inventory management. In response to these challenges, deep learning models have gained attention for visual recognition tasks. This review focuses on the capabilities of convolutional neural networks (CNNs), vision transformers (ViTs), and hybrid models combining both, with an emphasis on the Fashion MNIST benchmark. CNNs are widely used for extracting local spatial features, while ViTs excel at capturing global dependencies through self-attention mechanisms. Recent works have proposed hybrid architectures to leverage the strengths of both models, combining fine-grained feature extraction with broader contextual understanding. This paper offers a comprehensive survey of these hybrid approaches, evaluating their performance in terms of accuracy, scalability, and adaptability. Additionally, we highlight the importance of explainability in fashion classification models and discuss how hybrid solutions enhance model interpretability. Comparative findings suggest that hybrid models outperform individual architectures, making them highly effective for real-world fashion applications demanding both transparency and robust performance.
Bouzidi et al. (Fri,) studied this question.