A Unified Multiple Proxy Deep Metric Learning Framework Embedded With Distribution Optimization for Fine-Grained Ship Classification in Remote Sensing Images

Abstract

Improving ship classification performance in remote sensing imagery by deep metric learning (DML) is a newly emerging research topic and has good application prospects. From the perspective of the use of metric loss (classification loss and pairwise loss) and the way of proxy learning (a single proxy or multiple proxies), this study summarizes the existing DML methods into four representative frameworks and proposes a novel framework, namely, a unified multiple proxy deep metric learning framework embedded with distribution optimization (UMP+D). Specifically, the UMP+D not only unifies the combination of classification loss and pairwise loss into a single loss function containing only pairwise representation but also fuses it with multiple proxy learning. In addition, a distribution loss branch is embedded in the UMP+D to refine the distribution of samples in the feature embedding space to further tighten the intraclass samples and pull apart the interclass samples. Extensive experiments on two optical remote sensing datasets and one synthetic aperture radar dataset demonstrate that the proposed UMP+D framework outperforms the existing frameworks and achieves state-of-the-art performance.

Keywords

• Deep metric learning (DML)
• deep neural networks (DNN)
• fine-grained ship classification
• optical remote sensing (ORS)
• synthetic aperture radar (SAR)