Ph. D Student University of Washington Seattle, Washington, United States
Abstract: Here’s a revised version of your text:
In the development of Intelligent Transportation Systems (ITS), surveillance cameras are extensively deployed to ensure traffic safety and manage vehicle flow. However, these systems often encounter significant challenges when the camera lenses are stained or affected by raindrops. The interaction between light and lens stains can degrade image quality, reducing visibility and impacting the accuracy of vehicle detection and traffic monitoring.
This paper introduces an attention-based Generative Adversarial Network (GAN) model to restore traffic surveillance images degraded by lens stains during night-time conditions, caused by adverse weather such as rain and dust. Our model employs an innovative attention mechanism that identifies and prioritizes the regions most affected by stains, optimizing the restoration process. We also introduce a comprehensive dataset, consisting of both synthetic and real-world simulation images, designed to emulate the complex interaction between lens stains and varying lighting conditions commonly observed in night-time highway environments. This dataset enhances the model’s ability to simulate and rectify distortions effectively.
Experimental results show that our approach significantly improves the quality and reliability of the surveillance images, outperforming existing methods by delivering clearer and more accurate visual information under challenging conditions. Our findings demonstrate substantial improvements in image clarity and object detection. In particular, for vehicle detection tasks, mAP50 improved to 0.89 after restoration, compared to stained images. This underscores the model’s potential to enhance the robustness of traffic surveillance systems, especially in night-time scenarios.
Learning Objectives:
Attendees can expect to learn the following from this session:
Upon completion, participants will be able to describe the challenges associated with image degradation in traffic surveillance systems caused by lens stains and adverse weather conditions, like rain and dust.
Upon completion, participants will be able to explain the functioning and benefits of an attention-based Generative Adversarial Network (GAN) for restoring surveillance images affected by lens stains.
Upon completion, participants will be able to evaluate the effectiveness of the GAN model by comparing pre- and post-restoration vehicle detection accuracy using metrics like mAP50 in traffic surveillance.
