Publication

Abstract : Deep-metric learning using triplet loss has shown promise in remote sensing (RS) image retrieval by mapping similar images closer and dissimilar ones farther apart in embedding space. However, fixed-margin triplet loss struggles with intra-class and inter-class variability and often suffers from zero loss, where trivial triplets dominate training, causing ineffective convergence. To address this, we propose a novel hardness-aware adaptive margin triplet loss (HA2TL-CBIR) that dynamically adjusts the margin based on sample difficulty. By measuring the distance between positive and negative pairs, our method prioritizes hard triplets and reduces the impact of trivial ones’. This adaptive margin strategy overcomes zero-loss issues and enhances the learning of robust, discriminative features. Experiments on benchmark RS datasets PatternNet and NWPU-RESISC45 demonstrate the effectiveness of our approach in handling the growing complexity and size of RS image archives, improving retrieval performance significantly.