Learning-based methods for single image haze removal

Abstract

Haze is an atmospheric phenomenon where turbid media obscure the scenes. Haze reduces the visibility of the scenes and reduces the reliability of outdoor surveillance systems. Under severe hazy weather conditions, the aerosols scatters or sometimes completely stops the light rays from reaching the camera sensor. Thus, outdoor captured photos tend to be hazy in inclement weather and have low visibility. Color cast of captured photos in such inclement weather evidently also depend on the size of the aerosols and its properties. The major challenges needs to tackle in the field of image de-hazing are: low-visibility, color imbalance, image capturing medium, unavailability of real-world training data etc. This work mainly focuses on analyzing and designing different modalities for image de-hazing in the context of providing the solution to the above-mentioned challenges. The significant contribution of this work is in: 1) proposing a novel scene transmission map estimation method, 2) proposing a dense haze removal approach, 3) proposing a novel varicolored image de-hazing approach which is applicable for hazy images captured in different weather conditions 4) proposing an underwater image de-hazing approach and 5) proposing un-paired training network for image de-hazing. Accurate estimation of scene transmission map is a key to recover the haze-free image from input hazy image. In this work, a convolution neural network based approach is proposed for scene transmission map estimation. The contribution of the work lies in the haze relevant feature extraction from RGB and YCbCr color spaces of input hazy image and a novel feature fusion approach. Another contribution towards the image de-hazing is made by proposing an end-to-end deep network which is trained adversarially for dense haze removal. Along with the visibility improvement, restoration of color balance is also equally challenging problem in image de-hazing. In this work, we propose a varicolored image de-hazing network which restores the color balance in a given varicolored hazy image and recovers the haze-free image. Also, a large-scale synthetic varicolored hazy image database is generated to train the network for varicolored image de-hazing. Also, we have proposed an underwater image de-hazing approach which recovers perceptually pleasant images by improving the visibility and color balance in input underwater image. In general, a major hurdle to train a convolution neural network for image dehazing is the unavailability of large-scale real-world hazy, and corresponding hazefree image (i.e. paired data). Thus, in this work, an end-to-end network is proposed which is trained in an unpaired manner to resolve the unavailability of paired training data. The proposed image de-hazing approaches are evaluated on the current stateof- the-art databases such as D-Hazy, SOTS, HazeRD, NTIRE-2018, NTIRE-2019, RESIDE and set of real-world hazy images. Also, two new datasets are proposed in this work namely outdoor hazy image (OHI) dataset and synthetic varicolored hazy image (VHI) dataset. Standard quantitative evaluation parameters such as SSIM, PSNR, CIEDE2000 are used to evaluate the proposed de-hazing approaches.