Deep learning architectures for multiclass clothing recognition as the semantic core of automated virtual try-on systems

This article examines and substantiates the choice of deep learning architectures for multiclass clothing classification integrated into virtual try-on (VTO) systems. Systematically compared ResNet-50, EfficientNet-B4, and Vision Transformer (ViT-B/16) on DeepFashion2 and ModaNet datasets. ViT-B/16 achieved the highest accuracy of 92.4% Top-1 on DeepFashion2 and 88.9% on ModaNet, demonstrating an average cross-dataset accuracy drop of 3.9 percentage points, the smallest among evaluated models. Preliminary U²-Net segmentation statistically significantly improved macro-F₁ for all architectures (p < 0.001), with an average gain of 3.2 percentage points and reduction of the studio-to-street domain gap from 11 to 6 percentage points. EfficientNet-B4 provided the optimal accuracy-to-latency ratio, achieving 87% Top-1 accuracy at 60 FPS on consumer hardware (RTX 3060), while ViT-B/16 required optimization to maintain 45 FPS. The recommended strategy for industrial VTO systems combines U²-Net segmentation with architecture selection based on target platform capabilities, balancing visual fidelity and computational efficiency.