Sketch-based Image Retrieval using Unsupervised Learning

Sketch-Based Image Retrieval (SBIR) enables users to query image databases using freehand sketches, offering an intuitive alternative when textual input is ambiguous or unavailable. However, the practical deployment of SBIR systems is impeded by the considerable domain gap between abstract sketches and detailed natural images, the limited availability of annotated sketch-photo pairs, and the high variability inherent in human drawing styles. Recent advances in unsupervised learning have sought to address these limitations by leveraging unpaired and unlabeled data to learn transferable feature representations. This paper presents and evaluates contemporary unsupervised SBIR techniques, including instance-level contrastive learning, transformer-based sketch encoders, and style-invariant frameworks aimed at improving robustness to intra-class variations and noise. Furthermore, data-free and zero-shot retrieval approaches facilitate retrieval across unseen categories without requiring manual annotations. Additional strategies such as deep manifold alignment and sketch-to-photo synthesis further improve cross-modal alignment. Experimental results on benchmark datasets demonstrate the effectiveness of these methods in real-world, noisy, and unpaired scenarios. The review also explores future research directions, including multi-modal composite queries and meta-learning strategies, which aim to enable scalable and generalizable SBIR systems. Overall, unsupervised learning remains a central component in advancing SBIR by minimizing reliance on labeled data while enhancing retrieval performance.