Research On Deep Learning Models For Visibility Recovery From Single Hazy Image And Their Evaluations

In recent years,the pollution of haze amomg the world is becoming more intensified,and the problems of haze governance have aroused widespread concern from all walks of life.Haze and the decrease of visibility are not only harmful for the physical and mental health of human,but also greatly dangerous to residents' production activities.Atmospheric visibility plays a significant role in road traffic,particularly unpredictable agglomerate haze,which brings great danger to traffic participant.Consequently,improving image visibility by use of low-cost software becomes a promising choice in intelligent transportation systems.Aiming at restoration and evaluation of image visibility in hazy weather,this paper proposes following research points:1.Although extensive efforts have been made for dehazing,majorities of them heavily depend on accurate estimation of transmittance maps and also rely on the atmospheric scattering model in which complex parameter calculation is involved.Addressing those problems,this paper constructs an end-to-end supervised dehazing model for single hazy image.Specifically,an auto-encoder is exploited as a generator to restore the dehazed image in an attempt of preserving more image details.Moreover,the dehazing performance of the auto-encoder can be largely boosted via our advocated dual principles of discriminativeness,which is formed by re-exploring the classical dark channel prior as well as exploiting generative adversarial learning.The objective evaluation shows that the proposed approach performs competitively with state-of-the-art approaches.However,it outperforms them in terms of the visibility restoration especially as for the densely hazy images.While,it should be noted that the applicability of supervised deep dehazing methods is inherently problematic for practice,restricted by the artificially simulated training data pairs.2.To overcome the difficulties in collecting paired training data for supervised learning and also the difficulties in training Cycle-GAN-based unsupervised learning methods,a new unsupervised deep dehazing method is proposed by joint use of atmospheric scattering model and guided filtering for inversely generating a hazy image.This approach introduces an adversarial learning mechanism in the domain of hazy images,so as to better avoid the need of clear images.The experimental results demonstrate that new approach achieves a stronger dehazing ability than other unsupervised algorithms in terms of visibility restoration.While,the phenomenon of possible color shift is also observed in the dehazed images.3.In view of the inconsistency between objective evaluation and human visual evaluation,this paper additionally bases on the high-level visual task,i.e.,object detection,for evaluating various dehazing algorithms' practical performance.Experimental results show that performance evaluation by uses of object detection is more equitable and objective,which better facilitates the popularization and spread of promising dehazing algorithms.However,it is found that detection accuracy assisted by current dehazhing methods is far below the industry's requirements.