Image visual similarity with deep learning: application to a fashion ecommerce company

Abstract

Deep learning is a very trendy topic now, showing high accuracy in image based systems that can go from image segmentation to object detection and image retrieval. Because of this, multiple researchers and companies have been building and sharing work in the community, including pre-trained convolutional neural networks, available for public use. This work follows the trend and delivers an experimental study using deep learning for building a visually similar image retrieval application, comparing three different convolutional neural architectures for feature extraction and six distance indexes for similarity calculation in a real-world image retrieval problem, using real data from a fashion e-commerce platform from Morocco. After testing all the different combinations, we can conclude that for this dataset, Vgg19 combined with a correlation coefficient for similarity calculation is the tuple that best maximizes the similarity between a search image and its retrieved neighbors.