Cycle-Dehaze: Enhanced CycleGAN for Single Image Dehazing

Abstract

In this paper, we present an end-to-end network, called Cycle-Dehaze, for single image dehazing problem, which does not require pairs of hazy and corresponding ground truth images for training. That is, we train the network by feeding clean and hazy images in an unpaired manner. Moreover, the proposed approach does not rely on estimation of the atmospheric scattering model parameters. Our method enhances CycleGAN formulation by combining cycle-consistency and perceptual losses in order to improve the quality of textural information recovery and generate visually better haze-free images. Typically, deep learning models for dehazing take low resolution images as input and produce low resolution outputs. However, in the NTIRE 2018 challenge on single image dehazing, high resolution images were provided. Therefore, we apply bicubic downscaling. After obtaining low-resolution outputs from the network, we utilize the Laplacian pyramid to upscale the output images to the original resolution. We conduct experiments on NYU-Depth, I-HAZE, and O-HAZE datasets. Extensive experiments demonstrate that the proposed approach improves CycleGAN method both quantitatively and qualitatively.

Single image dehazing using deep neural networks

“…In addition to removal of as much haze as possible from images, we also aim to maintain two other critical characteristics including process speed and retention of fine detail. When comparing our method to Generative methods such as [9] while in some cases the GAN based method removes more haze, it is not able to return fine detail that an atmospheric scatter model based method can. This can be seen in Figure 2.…”

“…This is a departure from previous work, which have used the discriminator output to effectively exclude or keep already generated images to form a dataset for a later training process. We are broadly differentiated from Generative Adversarial Networks such as [22] and [9] as we are attempting to use a determinative model [15] to recover pixel based information from the source image as opposed to generating new data from learned parameters.…”

Single-Image Dehazing on Aerial Imagery Using Convolutional Neural Networks

“…Almost all of the single-image dehazing algorithms (Li et al, 2017, Li et al, 2018, Engin et al, 2018, Zhang et al, 2018, Min et al, 2019 use this model to reconstruct haze-free image by estimating the transmission map, airlight or combination of them.…”

“…Some of these images together with their dehazed versions have been illustrated in Figure 6. We use the Peak Signal To Noise Ratio (PSNR) and Structural Similarity (SSim) indicators as widely used metrics for dehazing performance evaluation (Li et al, 2017, Li et al, 2018, Engin et al, 2018, Zhang et al, 2018, Min et al, 2019. For the sample dehazed images, the PSNR and SSim comparisons are mentioned in captions of Figure 6.…”

“…The authors also added a percep-tual regularization based on the VGG network's feature-maps and applied L1 regularized gradient prior in order to avoid artifacts and color distortions in the dehazed images. A proposed method based on CycleGAN (Engin et al, 2018) uses two generators and two discriminators in the network to add or remove haze to the images. This method further takes advantage of cycle-consistency and perceptual losses to refine texture information and improve the final results.…”