Deep Pairwise Learning for Clothing Style Matching with DenseNet and Siamese Models

This study explores the use of a Siamese Neural Network (SNN) architecture combined with a DenseNet-121 model to predict outfit compatibility in the fashion domain. Trained on the Polyvore Outfits dataset, which includes categories such as tops, bottoms, shoes, and outerwear, so the model identifies compatible clothing items by learning and comparing detailed visual features like color and style. The outcomes of the training on the training set are confirmed by the experimental results, where an accuracy of 75.50% and a loss of 0.6850 were reached. However, the overfitting of the model demonstrated low generalization ability on the test data, with an accuracy of 72.45% and an increase in loss to 0.7358. This difference in performance between the training and test sets is also depicted by the precision and recall metrics. The results of the study illustrated a good ability of SNN to identify compatibility patterns but indicated low generalization skill among different outfits. This project provides a foundation for further optimization of deep learning models in fashion compatibility prediction.