Enhanced Bird Detection from Low-Resolution Aerial Image Using Deep Neural Networks

Abstract

Bird detection in LR images is essential for the applications of unmanned aerial vehicles. It is still a challenging task because traditional discriminative features in high-resolution (HR) usually disappear in low-resolution (LR) images. Although recent advances in single image super-resolution (SISR) and object detection algorithms have offered unprecedented potential for computer-automated reconstructing LR images and detecting various objects, these algorithms are mainly evaluated using synthetic datasets. It is unclear how these algorithms would perform on bird images acquired in the wild and how we could gauge the progress in the real-time bird detection. This paper presents a novel bird detection framework in LR aerial images using deep neural networks (DNN). We collect a dataset named BIRD-50 and a public dataset named CUB-200 of real bird images with different scale low-resolutions. Using these datasets, we introduce a novel DNN based framework for bird detection in reconstructed HR images, which exploits the mapping function from LR to HR aerial image and detects the birds by the state-of-the-art object feature extraction and localization methods. By systematically analyzing the influence of the resolution reduction on the bird detection, the experimental results indicate that our approach has produced significantly improved detection precision for bird detection by the inclusion of SISR algorithms.