A generative adversarial network model for visual monitoring of airport airside environment

• Main Author: Soo, Reine Jie Yi
• Other Authors: Sameer Alam
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Aeronautical engineering::Communication methods and equipment
• Online Access: https://hdl.handle.net/10356/158228
• Institution: Nanyang Technological University

Air Traffic Controllers (ATCOs) are responsible for managing air traffic flow and demands in the airspace. They rely on digital technologies such as PTZ cameras to aid in detecting and recognising faraway or small objects in the airport-airside environment. However, the farther the cameras zoom, the greater the drop in resolution of the image seen. Low resolution image hinders ATCOs from accurately determining the objects seen which can influence their situational awareness of the environment and decisions in managing the movement of aircraft. Single Image Super Resolution (SISR) is a technology that can improve the resolution of low-resolution images using digital capabilities. Whereas, Generative Adversarial Networks (GAN) is a deep learning model that can produce synthetic but photorealistic images from a pool of images. Several SISR architectures involving Generative Adversarial Networks (GAN) were explored and implemented to produce synthetic, photorealistic, high-resolution images that is an improvement to low-resolution real-world images. Among the architectures, the Real-ESRGAN architecture is proposed for SISR on images seen in PTZ cameras due to its stability in training, adaptability to real-world images and good balance between high resolution, sharpness and fineness in image detail. Moreover, in terms of quantitative comparison, the Real-ESRGAN model achieved consistently high Peak-Signal-to-Noise-Ratio (PSNR) and Structural Similarity (SSIM) among the chosen SISR models across the datasets used. PSNR and SSIM are common quantitative performance metrics for evaluating the performance of SISR models. Future works can be explored to push the limits of existing Real-ESRGAN architecture to attain higher resolution and quality in the image generated.