Haze removal from an image or a video via generative adversarial networks

• 主要作者: Chen, Zhong Jiang
• 其他作者: Loke Yuan Ren
• 格式: Final Year Project
• 語言: English
• 出版: Nanyang Technological University 2024
• 主題: Computer and Information Science; Haze removal; Generative adversarial networks
• 在線閱讀: https://hdl.handle.net/10356/181155
• 機構: Nanyang Technological University

Low visibility caused by haze and fog is one of the major reasons for traffic and aviation accidents. This paper introduces a more easy-to-access solution to remove haze from a single image, video, and live-streaming. My approach uses a modified conditional Generative Adversarial Network (cGAN) with a DenseNet-121 architecture to efficiently dehaze visual inputs. Unlike models that use Tiramisu [5] or depend on two-step pipelines, The modified model ensures the accuracy of structure and clarity of the visual by removing haze by optimizing the generator-discriminator interaction within the GAN framework. The effectiveness of the modified model is demonstrated through a comprehensive experiment on synthetic and real-world data, obtaining competitive results in PSNR, SSIM, and subjective quality measures. This system aims to improve visibility in live-streaming scenarios, such as for vehicles and aircraft, potentially reducing the probability of accidents under low-visibility conditions.