Interactive and light approach for near duplicate retrieval in multiple contexts

Samuel Lozachmeur, Thierry Urruty, Philippe Carré

Abstract: The increasing use of images in various industrial contexts, such as marketing, media, research, education, and documentation, has led to a constant growth in the size of image databases. As a result, the task of image retrieval has become even more challenging. To address this issue, we present a study on image retrieval in an industrial context, conducted in collaboration between a research laboratory and a company that is centered on Digital Asset Management (DAM). We propose an interactive and lightweight workflow for near-duplicate retrieval, using deep feature extraction and re-ranking techniques. Our first contribution focus on an unsupervised indexing approach for near-duplicate detection which achieves state-of-the-art performance on benchmark datasets. Then, we study the joint impact of different similarity metrics and deep convolutional neural network models on retrieval performance. We then create a new re-ranking distribution based on the chosen embedding and perform binary classification to refine near-duplicate detection. Lastly, we enable efficient user interactivity in selecting and visualizing near-duplicates, with respect to individual user preferences and context requirements.