Image retrieval systems based on embeddings for trustworthy AI

Abstract

Embedding-based image retrieval systems have become essential for applications that require high accuracy, computational efficiency, and adaptability, particularly in fields like healthcare and agriculture. These systems utilize deep learning models to generate high-dimensional feature embeddings, enabling scalable and precise image retrieval. This study evaluates two distinct retrieval pipelines: the Domain-Specific Retrieval Pipeline (DSRP) and the Pre-trained Feature Extraction Pipeline (PMRP). The comparative analysis, conducted across multiple datasets including medical imaging, agricultural analysis, and general object retrieval, highlights the advantages and trade-offs between these approaches.

The results demonstrate that PMRP achieves consistently higher retrieval accuracy across all datasets, with near-perfect performance in tasks like leaf disease detection and general object retrieval. In contrast, DSRP, while benefiting from domain-specific adaptation, does not surpass PMRP in overall performance and exhibits variability depending on dataset characteristics. Despite its domain- tuned architecture, DSRP showed increased computational complexity without a proportional increase in accuracy. Additionally, PMRP maintains more stable inference times, making it a preferable option for real-time applications.

Beyond retrieval accuracy, this study explores challenges related to dataset diversity, bias mitigation, explainability, and adversarial robustness. The findings underscore the importance of balancing retrieval precision with computational feasibility, particularly when deploying systems in high-stakes environments. Future research should focus on enhancing dataset representation, developing fairness-aware learning mechanisms, improving model transparency through explainable AI techniques, and strengthening adversarial defenses to ensure the security and reliability of embedding-based retrieval systems. By addressing these challenges, such systems can evolve into more efficient, interpretable, and ethically responsible solutions for real-world applications across various domains.