Saliency-Based Image Classification: A Survey of Methods, Applications, and Open Challenges

Saliency, the identification of the image regions that matter most for a visual decision, has evolved from a hand-crafted model of human attention into a general-purpose signal that is used to supervise, augment, compress, and explain modern image classifiers. This survey reviews recent work from 2024 to 2026 on saliency-based image classification and organises it into six categories: (A) saliency-guided data augmentation and mixup, (B) saliency-guided training objectives, (C) saliency-guided input and token selection for efficient inference, (D) saliency-guided feature mining, (E) saliency as explanation and its evaluation, and (F) robustness, trust, and security. For each category we summarise representative methods, the reported empirical gains, and the assumptions on which these gains rest. Across studies, we observe that (i) human or human-like saliency is a useful training signal when labelled data are scarce, (ii) Vision Transformer (ViT) attention has become the dominant saliency source for efficiency-oriented methods, and (iii) the faithfulness of saliency maps remains contested, which limits their use as clinical or safety evidence. We close with a discussion of open problems, including standardised evaluation, the cost of collecting human saliency, and the interaction between saliency and robustness. All quantitative results are reported as stated by the original authors and are not directly comparable across studies.

Nine formats, generated from this paper’s own record. Pick the one your journal or department asks for, or download the entry straight into your reference manager.

Counts exclude identified crawler traffic and repeat hits from the same reader within 30 minutes. Downloads reached directly through a search engine or an existing citation link are not included, so the figure is a conservative count of activity through this site.

© 2026 - Authors hold the copyright of this article. This work is licensed under a Creative Commons Attribution 4.0 International License. and The Open Definition. This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (CC BY 4.0). 🛡️ Disclaimer: The content, data, and findings in this article are based on the authors’ research and have been peer-reviewed for academic purposes only. Readers are advised to verify all information before practical or commercial use. The journal and its editorial board are not liable for any errors, losses, or consequences arising from its use.

JMVRA is a scholarly open access, peer-reviewed and refereed journal with an impact factor of 11.25, calculated from Google Scholar citation data. It is a multidisciplinary, monthly, low-cost and transparent peer review publication aligned with UGC’s Suggested Parameters for Choosing Peer-Reviewed Journals.

Publishing with JMVRA includes indexing through Google Scholar, Crossref, OpenAlex and other verified discovery services, an optional Crossref DOI for each article, citation generation tools, and full open access visibility — so the work reaches readers and accrues citations rather than sitting behind a paywall.

Journal of Modern and Versatile Research Approach exists to advance applied, theoretical and experimental research across academic and professional fields, connecting researchers, developers, academicians, engineers and practitioners through peer-reviewed publication and scientific collaboration.

Indexing & Discovery — verified records

JMVRA operates a strict double-blind peer review system managed by 3000+ expert reviewers, so every paper is assessed ethically, without bias, and on its merits. The review is transparent: you can obtain the report itself, not just the verdict.

Journal Type: JMVRA is an international open-access journal offering low-cost publication, transparent peer review, Crossref DOI and multidisciplinary research visibility, aligned with UGC’s Suggested Parameters for Choosing Peer-Reviewed Journals.