Enhancing Fine-Grained Image Retrieval with Advanced Embedding and Loss Functions
Mustafa Keskin, Alp A. Yalman, Enis Teper, Sinan Keçeci, Emre Rençberoğlu · 2025
This paper explores the optimization of similarity retrieval systems through the fine-tuning of embeddings using advanced loss functions, focusing on the Stanford Cars dataset. By employing Triplet Loss with hard and semi-hard negative sampling, alongside the ArcFace loss, we significantly enhance retrieval accuracy, demonstrated by increased Retrieval R-Precision. Our methodology includes robust preprocessing and data augmentation, utilizing a ResNet-based architecture to extract high-level image features. An efficient retrieval pipeline is established through the Hierarchical Navigable Small World (HNSW) algorithm, ensuring real-time performance. The results indicate substantial improvements in capturing fine-grained visual similarities, with potential applications across autonomous vehicles, retail, and beyond. Future work will involve extending the approach to diverse datasets, exploring advanced architectures, and optimizing for real-time efficiency.