CANDY: Conditional Adversarial Networks based End-to-End System for Single Image Haze Removal

Abstract

Single image haze removal is a very challenging and ill-posed problem. The existing haze removal methods in literature, including the recently introduced deep learning methods, model the problem of haze removal as that of estimating intermediate parameters, viz., scene transmission map and atmospheric light. These are used to compute the haze-free image from the hazy input image. Such an approach only focuses on accurate estimation of intermediate parameters, while the aesthetic quality of the haze-free image is unaccounted for in the optimization framework. Thus, errors in the estimation of intermediate parameters often lead to generation of inferior quality haze-free images. In this paper, we present CANDY (Conditional Adversarial Networks based Dehazing of hazY images), a fully end-to-end model which directly generates a clean haze-free image from a hazy input image. CANDY also incorporates the visual quality of haze-free image into the optimization function; thus, generating a superior quality haze-free image. To the best of our knowledge, this is the first work in literature to propose a fully end-to-end model for single image haze removal. Also, this is the first work to explore the newly introduced concept of generative adversarial networks for the problem of single image haze removal. The proposed model CANDY was trained on a synthetically created haze image dataset, while evaluation was performed on challenging synthetic as well as real haze image datasets. The extensive evaluation and comparison results of CANDY reveal that it significantly outperforms existing state-of-theart haze removal methods in literature, both quantitatively as well as qualitatively.Both the authors are affiliated with Samsung R&D Institute India -Bangalore

Unsupervised Dark-Channel Attention-Guided CycleGAN for Single-Image Dehazing

“…Isola et al [20] completed the style transformation of paired images through the conditional generation antagonism network (CGAN) [20]. GANs [7] are also often used in dehazing [21,22]. Visual Geometry Group Network (VGG) features was introduced by Li et al [23,24] introduced, and the L1-regularized gradient of CGAN [20] was used for image dehazing.…”

Robust Computing for Machine Learning-Based Systems

“…• Gradient Sign-based Adversarial AƩacks [28,25,43] • OpƟmizaƟon-based Adversarial aƩacks [6,54] • Backdooring AƩacks [15] • Pruning-based Defenses [15] • Preprocessing-based Defenses [26,27,3,55] • GAN-based Defenses [52,9,63,67] • Methodology for Building Resilient Hardware [18] • Error-Resilience Analysis [18,17] • Fault-Aware Pruning (FAP) [66] • Fault-Aware Pruning + Training (FAP+T) [66] • Timing Error-Drop (TE-Drop) [64] • StaƟc Voltage Underscaling (ThVolt-StaƟc) [64] • Per-layer Voltage Underscaling (ThVolt-Dynamic) [64] Fig . 3 Overview of the works discussed in this chapter for addressing reliability and security vulnerabilities of deep learning-based systems 3.…”

Cycle-Dehaze: Enhanced CycleGAN for Single Image Dehazing

“…depth information or known 3D model of the scene. Single image dehazing approaches are divided into prior information-based methods [3,5,12,16,20,32] and learning based methods [10,24,30,34,35]. Prior information-based methods are mainly based on the parameter estimation of atmospheric scattering model by utilizing the priors, such as dark channel priors [16], color attenuation prior [38], haze-line prior [8,9].…”

“…Even though most of the deep learning approaches use the estimation of intermediate parameters, e.g. transmis-sion map and atmospheric light [10,24], there are also other approaches based on generative adversarial networks (GANs), which build a model without benefiting from these intermediate parameters [30].…”

“…Their major goal is the generation of fake images indistinguishable from the original images on the targeted domain. By utilizing GANs, there exist state-of-the-art methods [30,35] for single image dehazing, which require hazy input image and its ground truth in a paired manner. Recently, the need of paired data is removed after the cycle-consistency loss has been proposed by CycleGAN [37] for image-to-image translation.…”