Deep learning-based image dehazing across different color spaces

Abstract

A difficult problem in computer vision is image dehazing. Since haze alters the color of objects in an image and weakens the intensity of certain light wavelengths, the color becomes a crucial factor to consider when we are trying to remove haze. Primarily, significant performance is achieved using deep learning-based computer vision in RGB, HSV and LAB color models. This study examines the effects of deep convolu􀆟onal neural networks based on GAN architectures on the elimina􀆟on of haze in various color spaces. The GAN architecture has generator and discriminator com- ponents which work in an adversarial manner, constantly improving and refining the dehazing process. The generator aims to generate more realis􀆟c and visually appealing haze-free images, while the discriminator strives to become more adept at dis􀆟nguishing real and generated images. A􀅌er training and tes􀆟ng in twelve different color spaces, the six deep-learning models show that, for most cases, XYZ, RGB, RGBCIE, and YDbDr consistently beat other color models. Due to differences in pixel intensity informa􀆟on across different color models, the color space transla􀆟on causes change in the image’s color characteris􀆟cs under foggy situa􀆟ons. In this study, the previously described idea is used to remove haze from foggy images using deep learning networks. The outcomes show a range of responses, with some networks outperforming RGB, HSV, and LAB in terms of SSIM and PSNR. The dehazing performance in RGB, HED, HSV, LAB, RGBCIE, XYZ, YUV, YCbCr, YDbDr, YIQ, YPbPr, and YUV color spaces is discussed in this paper. The goal of the study is to demonstrate how deep-learning models perform in elimina􀆟ng haze in various color spaces.