Deep Multi-scale Residual Learning for Single Image De-Hazing

Images play an important role in various applications such as computer vision, scenario analysis, object detection, advanced driver assistance system (ADAS) etc. Usually images that are captured from outdoor get degraded due to the presence of turbid media and causes the outdoor vision system to provide deficient performance. The turbid media can be rain, haze, smog, fog, snow etc. With the help of developing technologies, it has become easier for everyone to recover such kind of distorted images. In order to overcome the limitations of existing models a new variant called deep multi-scale residual network has been proposed. The multi-scale deep residual learning network along with the simplified convolutional neural network map the hazy image to its corresponding haze-free image there by providing dehazed output image. The basic building block of the multi-scale residual network, passes the output from the previous layer and is fed to next layer. So that the possible color distortion in the recovered image would be avoided. The simplified CNN designed with the help of generic agnostic model helps in enhancing the haze-free image. Qualitative and quantitative analysis is performed. A simplified architecture with less parameters is designed and the model shows an excellent performance compared to state-of-art methods.

I. Introduction

Different weather conditions, such as haze, fog, mist, smoke, fumes, dust, rain as well as snow, will cause unfavourable visual effects in images or videos. Such artifacts may significantly degrade the performances of several outdoor vision systems, such as event detection and understanding, object detection, tracking, and recognition, and scene analysis and classification, even for outdoor surveillance or ADAS (advanced driver assistance systems). Moreover, hazing artifacts have also shown significant influence on the performance of algorithms that requires high quality images as input and also affect the storage efficiency for digital visual data. The work focuses on removing haze (or de-hazing) from a single image. Based on the survey conducted on recent works in single image haze removal using deep learning, most of them were not able to capture the intrinsic characteristics of hazy images and by using these approaches the obtained haze-free images tends to have less image quality. To obtain better image de-hazing results compared to existing methods a novel and enhanced method has been developed.