<italic>DRL</italic>-GAN: Dual-Stream Representation Learning GAN for Low-Resolution Image Classification in UAV Applications

Abstract

Identifying tiny objects from extremely low-resolution (LR) unmanned-aerial-vehicle-based remote sensing images is generally considered as a very challenging task, because of very limited information in the object areas. In recent years, there have been very limited attempts to approach this problem. These attempts intend to deal with LR image classification by enhancing either the poor image quality or image representations. In this article, we argue that the performance improvement in LR image classification is affected by the inconsistency of the information loss and learning priority on low-frequency (LF) components and high-frequency (HF) components. To address this LF-HF inconsistency problem, we propose a dual-stream representation learning generative adversarial network (DRL-GAN). The core idea is to produce enhanced image representations optimal for LR recognition by simultaneously recovering the missing information in LF and HF components, respectively, under the guidance of high-resolution (HR) images. We evaluate the performance of DRL-GAN on the challenging task of LR image classification. A comparison of the experimental results on the LR benchmark, namely HRSC and CIFAR-10, and our newly collected `WIDER-SHIP' dataset demonstrates the effectiveness of our DRL-GAN, which significantly improves the classification performance, with up to 10% gain on average.

Keywords

• Convolutional neural networks (CNNs)
• generative adversarial networks
• low-resolution (LR) image classification
• representation learning
• unmanned aerial vehicle (UAV)-based remote sensing