Data Science Master's Degree Graduation Thesis
Fashion industry, one of the world’s largest industries, is facing a constantly increasing problem: with new technologies and social media, fashion trends from all over the world constantly influence customers taste, which is becoming always more ever-changing. Fashion brands, with their stylists and designers, need to continue to produce successful items by creating always new, trendy and up-to-date items to satisfy the taste of their customers. Computer vision, generative models and deep learning applied in fashion domain can become good allies to stylists and designers creativity giving them a new source of inspiration. Indeed, the aim of this work is to develop a system able to create new and realistic fashion items starting from existing clothes images. Thus, two different tasks are developed: a preliminary one of fashion item detection to crop images and make them more suitable for the second task of image generation. To carry out these two tasks, state-of-the art models are used: Faster R-CNN object detector model, trained on DeepFashion dataset, and a custom DCGAN that generates new items starting from images of fashion brands’ runways collected ad hoc by scraping the social network Pinterest.