Deep Learning Enhanced UAV Imagery for Critical Infrastructure Protection

Abstract

Unmanned aerial vehicles (UAVs) have seen a significant increase in their commercial application because of various technological break-throughs. As UAVs are typically used in open environments for purposes such as military applications, surveillance, and delivery of commodities, they rely primarily on the visual signals recorded by the flying UAV. UAVs can be used for critical infrastructure protection where their surveillance capabilities are used for monitoring these sites. While carrying out such missions in open environments, visual degradation is an unavoidable concern. It has a negative impact on the performance and security of the system. We propose a deep-learning-based framework, called Aerialgan, to solve the visual degradation caused by haze in the atmosphere. We trained the proposed model using an adversarial training algorithm which is commonly known as the generative adversarial ntwork (GAN) and aims to enhance the hazy images collected by a UAV and generate a clean, haze-free image of the same scene. In addition, we present the Aerial Non-Homogeneous Hazy (ANHH) dataset, which contains over 66,000 pairs of hazy and ground truth aerial photos with realistic, non-homogeneous haze of various densities. We used performance metrics such as peak signal-to-noise ratio and structural similarity index to evaluate our model on ANHH and compare it with contemporary state-of-the-art techniques in image dehazing. The proposed technique can be very useful in improving the reliability of surveillance where UAVs are used for critical infrastructure protection applications.