Content-based Image Retrieval using an Adaptive Triplet Learning Approach from Multi-label Remote Sensing Image Archives

Rapid expansion of remote sensing (RS) image archives necessitates the development of advanced content-based image retrieval (CBIR) methods to effectively organize and access high-dimensional visual data. Deep metric learning, particularly with triplet loss, has shown promising results in retrieval tasks by embedding semantically similar images closer in feature space. However, conventional fixed-margin triplet loss functions are inadequate for multi-label RS archives, where high intra-class variability and inter-class similarity is observed. These limitations often result in ineffective training due to the dominance of trivial triplets and the zero-loss problem. In this work, we propose a novel multi-label adaptive triplet learning approach for CBIR in RS archives. The proposed method introduces an adaptive margin approach that dynamically adjusts the triplet loss margin based on the hardness of each sample, computed as the difference between anchor-positive and anchor-negative distances. This adaptive strategy enables the model to focus on informative (hard) triplets and suppresses the influence of trivial ones, leading to more discriminative feature embeddings. We evaluate our proposed approach, on two challenging multi-label RS archives: MLRSNet and UCMerced. Experimental results demonstrate that the proposed method consistently outperforms fixed-margin triplet loss and several state-of-the-art CBIR approaches, particularly in retrieval accuracy and robustness. These findings confirm the effectiveness of adaptive margin learning in handling the complexities of multi-label RS image retrieval.

keywords: Content-based image retrieval (CBIR), Remote Sensing, Deep Metric Learning, Multi-Label Image Retrieval, Triplet Loss, Adaptative Margin