Image Dehazing by Joint Estimation of Transmittance and Airlight Using Bi-Directional Consistency Loss Minimized FCN

Very few of the existing image dehazing methods have laid stress on the accurate restoration of color from hazy images, although it is crucial for proper removal of haze. In this paper, we are proposing a Fully Convolutional Neural Network (FCN) based image dehazing method. We have designed a network that jointly estimates scene transmittance and airlight. The network is trained using a custom designed loss, called bi-directional consistency loss, that tries to minimize the error to reconstruct the hazy image from clear image and the clear image from hazy image. Apart from that, to minimize the dependence of the network on the scale of the training data, we have proposed to do both the training and inference in multiple levels. Quantitative and qualitative evaluations show, that the method works comparably with state-of-the-art image dehazing methods.

1. Introduction

Haze and fog reduces the visibility of outdoor scenes. For this reason, distinguishing objects from distance becomes difficult. Haze occurs when light falls on atmospheric particles and gets absorbed and scattered by them. This causes deterioration in the quality, particularly contrast, of the captured image. The strategy for eradicating the effect of haze from such degraded images is known as Image Dehazing (Fig. 1). Image dehazing is a tricky problem to solve due to the direct dependence of the haze density on the depth of objects. Diverse methods have been suggested to tackle the problem with impressive outcomes [23], [18], [15], [14]. Estimating the scene transmittance and environmental illumination has been established as the key to solve this problem. In recent times, single image dehazing has been receiving a lot of attention due to its practical significance. Due to the ill-posed nature of the problem, the methods mainly depend on statistical priors and physical cues. The recent success of Convolutional Neural Networks (CNN) in the field of computer vision [19], [10], [12] have inspired its use in image dehazing [8], [24], [20]. The main advantage of CNN is that they can learn features from data. This enables feature learning based-on the hazy image formation mechanism. Figure 1.

Hazy image and its dehazed version obtained by our method