An Image Similarity-based recommendation system for e-Commerce

In this thesis, we propose multiple approaches to accomplish the image similarity in order to present to the customers of a e-commerce the most suitable options based on the products that this person choose. The methods used in this thesis cover the classic processing of images, unsupervised deep learning models, bag of visual words, autoencoders and supervised siamese network that were trained based on the labeled data from the staff of the fashion company where the data come from. Given this logic having the representation of the images on embeddings we can calculate the ranking of the products from the most similar to the most different with cosine or euclidean distance. In the end, we compare the performance of each model given the baseline of the labeled data of the product and we can deliver a final statement of which one is the most convinient to the company. During this document you can find as well the preprocessing of the images of the products and the standarization to obtain comparable results.