Attribute Prediction and Garment Similarity Evaluation from Fashion Images

Abstract

This thesis presents a study of advanced garment classification and recommendation systems in hopes to improve online fashion retail experiences. Utilizing Convolutional Neural Networks (CNNs), we propose a method to classify an image into distinct categories and subcategories. Additionally, our approach, from an input image, is able to analyze a garment’s attributes such as pattern, fabric, and style. These findings enable nuanced garment similarity and compatibility analyses. Our methodology leverages the DeepFashion dataset, employing pre-trained ResNets in order to best handle the garment classification and attribute prediction task. The models demonstrate promising results in identifying garment categories, subcategories, and attributes, as well as in garment similarity analyses highlighting the potential of transforming fashion retail by making online shopping more personalized and efficient.