S-HAZE: Dataset Consisting of Real World Ground Truth and Hazy Images of Varying Haze Density with No-Sky,Little Sky and Large Sky Regions

Abstract

De-hazing images captured during inclement weather is highly necessary for proper visibility restoration. In recent years many researchers have proposed various de-hazing algorithms where they mostly portrayed de-hazing as a simple contrast enhancement problem. Proper evaluation of efficiencies of proposed algorithms is often not possible due to lack of sufficient datasets. To overcome this shortcoming, here we have introduced a new dataset S-HAZE, which consists of 60 Ground truth (GT) images with large sky, little sky and no-sky regions. Sky regions are also of different types like clear blue sky, blue sky with clouds, white sky, etc. The reasons behind focusing on sky regions of images while constructing the dataset are as follows: a. Sky regions play a crucial role in the atmospheric light evaluation, whose accurate estimation is highly essential for artifact free high quality image recovery. b. Most of the existing de-hazing algorithms are designed assuming the sky regions of haze free images to be white ideally but in reality this assumption is not always true, so proper restoration of scene color of sky regions in images are often not achieved by these algorithms. That's why to check the capability of true sky color restoration of these algorithms we have included images with different types of sky regions in this dataset. c. Performance of de-hazing algorithms vary according to the variation of haze density in images. So here we have included two hazy images corresponding to each GT image with varying haze density (light haze and dense haze) in the dataset to notice how the performance of algorithms vary with the haze density in same image. Hazy images in the dataset are taken in the presence of real haze created by artificial haze machines. Finally while conducting the quantitative analysis of output images obtained from different de-hazing algorithms using three quantitative parameters like Structural Similarity Index (SSIM), Haze improvement (HI) and Mean square error (MSE) we noticed another interesting finding that de-hazing algorithms giving desired maximum SSIM and minimum MSE values give poor HI value and vice versa. Hence, we concluded that portraying de-hazing as a mere contrast enhancement procedure is incorrect.