Abstract
Removing the haze in an image is a huge challenge due to the difficulty of accurate hazy image modeling. Although the atmospheric scattering model (ASM) is widely used to describe the formation of hazy images, it is hard to deal with the uneven haze image, once the ASM is restricted with the assumption that the atmosphere distributed homogeneously. This paper analyzes the haze imaging mechanism, then proposes an image-to-image architecture to handle the uneven image dehazing, in which a heterogeneous twin network (HT-Net) with two parallel sub-networks are constructed to establish the high dimensional nonlinear mapping model between the hazy and clean images. Consequently, the inhomogeneous haze is removed by the symmetric U-shape network with encoder-decoder structure, meanwhile, the other enhancement network extracts the high-frequency feature from the hazy image to compensate the edge and texture of the object. The effectiveness is validated by the experiments based on three real haze image datasets which depict the same visual content recorded in haze-free and hazy conditions, under the same illumination parameters. One professional uneven haze image dataset is found in the real environment and covers 190 types of scene and 21975 uneven images accordingly. This dataset includes thin, heavy, and uneven haze images. The other two benchmark datasets are I-HAZE and O-HAZE, respectively including 35 pairs of indoor real haze and haze-free (ground-truth) images and 45 different outdoor scenes. Extensive experimental results demonstrate that the proposed method in this paper can remove the haze and achieve superior performance over the other mentioned methods.
Highlights
Images captured by the devices often have been degraded by the kinds of inclement weather conditions, such as rain, snow, fog, sand storms
Aiming at the uneven haze image, this paper proposes a framework with two different structures called heterogeneous twin network (HT-Net), this framework can remove the thin, heavy, and uneven haze and enhance the loss of the edge information adaptively
2) COMPARISONS WITH OTHER METHODS To demonstrate the effectiveness of the proposed network, we perform the following three qualitative experiments on various hazy images data and one quantitative comparison on Peak Signal-toNoise Ratio (PSNR), Structural Similarity (SSIM), and Mean Square Error (MSE) with five existing dehazing methods: Dark channel prior (DCP) [8], DeHazeNet [13], multi-scale convolutional neural network (MSCNN) [14], AOD-Net [15] and FFA-Net [17]. 21450 images are been used to train the network and 525 images as a non-overlapping test data set
Summary
Images captured by the devices often have been degraded by the kinds of inclement weather conditions, such as rain, snow, fog, sand storms. Our approach mainly handles the uneven hazy image without the limitation that the atmosphere is homogeneous and the proposed network can supplement the detail information lost in the feature extraction by enhancement network to restore the high-pixel dehazed image, while our method is imperfect to a few images with the color closed to white in a large area. For these images, the output of HT-Net would be supersaturated in a mono-color. There exist plenty of hazy images collected indoors or outdoors, which are hard to ensure the atmosphere is homogeneous, and in some cases, the images have uneven haze exactly
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
