Monochromatic Image Dehazing Using Enhanced Feature Extraction Techniques in Deep Learning

Images photographed in foggy weather usually have poor visibility. To mitigate this problem researchers have come up with various image dehazing techniques. Now, more than ever, high-quality images that can be used to glean maximum information from autonomous systems are in high demand. This research work uses different Deep Learning (DL) architectures to draw out essential details from the picture and localize the information recovered to reduce the haze from the picture. The paper investigates to remove the hazes from the dehazed images using DL techniques. The first task of this proposed work attempts three pre-processing techniques namely, air light estimation, contextual regularization and boundary constraint. The second task of this work is to identify the suitable DL model to extract clear images from dehazed images. Evaluation metrics are PSNR value and SSIM value are used to estimate the values of dehazed images compared with clear images. Experimental results proves that AOD-Net outperforms good result with respect to PSNR value.

I. Introduction

A photograph taken in foggy conditions results in an image with low visibility. As depicted in Figure 1, the haze causes distant objects to lose contrast and blend in with their surroundings. The light reflected by these items is dimmed and diluted by the environment, and it interacts with light dispersed by various particles in the air before reaching the camera. Consequently, as these objects recede further from the camera, their colors fade and become more reminiscent of fog, with the degree of resemblance increasing with distance.