A Study on Performance Optimization of Ensemble-Based Clothing Image Classification Model Using Knowledge Distillation

다국어 초록 (Multilingual Abstract)

This study was conducted with the goal of improving the accuracy of clothing image classification and ensuring system efficiency. In particular, it focused on enhancing classification performance for visually similar and challenging clothing classes. To achieve this, models from the Swin Transformer, ConvNeXt, and EfficientNet families—pretrained on large-scale datasets—were fine-tuned for the clothing domain, and a re-labeled clothing image dataset was constructed and utilized. To overcome the limitations of individual models, a Soft Voting ensemble was constructed based on the selected backbone models. Experimental results demonstrated that the ensemble model significantly outperformed single models in both pattern and category classification tasks, achieving notable improvements in Macro-F1 score (approximately 4.8 percentage points for pattern classification and 1.2 percentage points for category classification). The ensemble also showed superior prediction capabilities for confusing classes. In addition, the study applied the Knowledge Distillation (KD) technique to maintain the high performance of the ensemble model while achieving model compression. As a result, the lightweight ConvNeXt-Tiny student model showed substantial gains in inference efficiency—improving inference speed by approximately 85% in category classification and 78% in pattern classification—while maintaining comparable classification performance. These experimental results demonstrate the potential applicability of the proposed approach to real-world online fashion retail systems, where both high classification accuracy and real-time inference speed are required.