A super-resolution enhancement of UAV images based on a convolutional neural network for mobile devices

The task of reconstructing high-resolution images given low-resolution images is called image super-resolution (SR). Within the wide variety of RS applications, it has a strong impact on the advancement of unmanned aerial vehicle (UAV) technologies. The camera is one of the most commonly used sensors in current UAVs, allowing complex information to be collected from the environments to be monitored. However, UAVs usually have limitations regarding their flight time. If one wants to monitor a large geographic area in a short time, flights must be performed at higher altitudes. This implies a loss of spatial resolution because cameras have limitations in terms of their optics and the size of the pixels. In recent years, SR techniques based on convolutional neural networks (CNNs) that are able to learn the correspondence between low-resolution images and their high-resolution counterparts have been developed. Taking advantage of advances in smartphones in terms of popularity and computing capabilities, the aim of this paper is to demonstrate an architecture for the SR of images captured by a UAV. The images are sent from a low-resolution camera in a UAV to a mobile device from which it is possible to obtain the corresponding SR image. A benchmark dataset of images is used for both quantitative and qualitative assessment.

A super-resolution enhancement of UAV images based on a convolutional neural network for mobile devices Articles