Enhancing Generalization in Sketch-Based Image Retrieval through Single and Multi-Source Domain Adaptation.

Abstract

This thesis addresses the critical challenge of generalization in deep neural networks, particularly within the context of Sketch- Based Image Retrieval (SBIR). A primary contribution is the de- velopment of a novel empirical framework to evaluate and bench- mark the generalization capacity of deep networks. This frame- work introduces metrics that quantify both model accuracy and the ability to handle data diversity, offering a practical approach to assess model performance on unseen data and identifying trade- offs crucial for effective model selection. Building on this generalization framework, the research proposes domain adaptation strategies specifically tailored for SBIR to bridge the significant gap between sketch and image domains . A single-source domain adaptation algorithm is introduced, uti- lizing canonical correlation analysis (CCA) alongside dictionary learning principles and sparse optimization techniques to facili- tate effective knowledge transfer from a source (e.g., images) to a target domain (e.g., sketches), even in few-shot scenarios . This approach is further extended to a multi-source domain adapta- tion algorithm, capable of integrating information from multi- ple diverse source domains to enhance robustness and adaptabil- ity. Computational efficiency is a key consideration, addressed through the use of low-rank matrix decomposition and online dictionary learning techniques. Overall, this work provides a comprehensive approach to enhanc- ing SBIR system performance by directly tackling generalization limitations through principled empirical assessment and efficient single- and multi-source domain adaptation methods. The find- ings contribute to both the understanding of generalization in deep learning and the development of practical, adaptable SBIR systems.