Comparative Review on Image Dehazing Methods Using Deep Learning Approach for Vision-Based Applications

Vision-based applications such as traffic management, public area surveillance, and environmental monitoring require clear and accurate visual data from cameras to provide better services to people. But foggy weather and traffic pollution causes haze, dust, or smog to degrade image quality, and reduce visibility of the target area, and becomes a challenge to the reliability of the automated application in urban areas. They face some additional problems such as color distortion, non-uniform illumination, low light, and dense or nonhomogeneous haze. The learning-based methods outperform traditional image dehazing methods in complex real-world problems and utilize large image datasets, despite their notable performance, learning-based models can have high computation and may require wide-ranging training datasets. This paper discussed prior-based and deep learning-based image dehazing models to test the RESIDE dataset with indoor and outdoor hazy images. Further, objective evaluation metrics have been discussed that are being used to evaluate the quality of state-of-the-art methods. These methodologies help us to set up vision-based applications for smart cities and have shown great potential to enhance image quality affected by haze and low light.