A Drone-Based Image Dataset Generation Methodology for Single Image Super- Resolution.

Abstract

The advancements in drone technologies, digital imaging, computer vision techniques, and the liberalized laws related to drone flying have opened up drone-based applications such as the delivery of supplies, search and rescue, aerial surveillance, and so on. The drones, especially the nano/micro/small drones, may be mounted with only low-resolution camera(s) due to their maximum takeoff weight limitations. The low-resolution images generated by the cameras, if used for landing, can result in faulty detection unless the photos are taken from a very close distance to the point of interest. Detection and recognition of the point(s) of interest as early as possible is required to ensure sufficient response time for safe maneuvering. Hence, the images are to be captured at greater heights or distances from the point(s) of interest, and obtaining the high-resolution images from the captured low-resolution images is crucial. The High Resolution (HR) and the Low Resolution (LR) image pairs for training super-resolution models in the works presented in literature are generated using two different cameras or the HR images are captured by the camera and LR images are generated by degrading the HR images. As both methods are not appropriate for small/ micro/nano category drones, we propose a novel method based on Ground Sampling Distance (GSD) to capture the LR and HR images. In this paper, we have presented the designed methodology for the creation of a dataset using drone-mounted cameras covering a broad spectrum of views of the target(s) suitable for training and testing of the Single Image Super-Resolution (SISR) models. We also present a methodology for selecting an appropriate target for imaging that enables the visual quality assessment of the developed super-resolution model.
Subjects
AERIAL surveillance; COMPUTER vision; SITUATIONAL awareness; HIGH resolution imaging; DEEP learning

Description