Parametric Analysis and Removal of Haze from Image by Convolutional Neural Network

Information extraction and processing using visual sensors is one of the main objectives in numerous vision-based applications like border surveillance, environmental or agricultural field monitoring, automatic cars, etc. One of the biggest challenges in getting the exact information is haziness, which can occur due to Fog/mist, Smoke, cloudy weather and many more. This research work proposes a parametric analysis and learning-based haziness removal technique for extracting more accurate information from the underlying scene. The degree of haziness (h) is estimated from the input images using statistical modelling combining the amount of light and transmission map. Subsequently, haziness is reduced significantly using deep learning. The proposed approach classifies input image streams into three different categories like clear, light hazy and dense hazy image depending upon the corresponding h values. The proposed approach is confirmed for the desired output using two publicly available datasets NYU and SOTS. The result shows the efficiency of haze removal in terms of PSNR and SSIM compared to state of the art research works.