Unsupervised affinity learning based on manifold analysis for image retrieval: A survey

• Main Author: Pereira-Ferrero, V. H. [UNESP]
• Publication Date: 2024
• Other Authors: Lewis, T. G., Valem, L. P. [UNESP], Ferrero, L. G.P., Pedronette, D. C.G. [UNESP], Latecki, L. J.
• Format: Other
• Language: eng
• Source: Repositório Institucional da UNESP
• Download full: http://dx.doi.org/10.1016/j.cosrev.2024.100657; https://hdl.handle.net/11449/309508
• topic: Affinity learning; Diffusion process; Image retrieval; Manifold learning; Multimedia retrieval; Ranking; Unsupervised

Despite the advances in machine learning techniques, similarity assessment among multimedia data remains a challenging task of broad interest in computer science. Substantial progress has been achieved in acquiring meaningful data representations, but how to compare them, plays a pivotal role in machine learning and retrieval tasks. Traditional pairwise measures are widely used, yet unsupervised affinity learning approaches have emerged as a valuable solution for enhancing retrieval effectiveness. These methods leverage the dataset manifold to encode contextual information, refining initial similarity/dissimilarity measures through post-processing. In other words, measuring the similarity between data objects within the context of other data objects is often more effective. This survey provides a comprehensive discussion about unsupervised post-processing methods, addressing the historical development and proposing an organization of the area, with a specific emphasis on image retrieval. A systematic review was conducted contributing to a formal understanding of the field. Additionally, an experimental study is presented to evaluate the potential of such methods in improving retrieval results, focusing on recent features extracted from Convolutional Neural Networks (CNNs) and Transformer models, in 8 distinct datasets, and over 329.877 images analyzed. State-of-the-art comparison for Flowers, Corel5k, and ALOI datasets, the Rank Flow Embedding method outperformed all state-of-art approaches, achieving 99.65%, 96.79%, and 97.73%, respectively.