To describe the composition of hazy images, McCartney [6] first proposed the use of an atmospheric scattering model, which was developed by Narasimhan and Nayer [7][8]. Equation (1) describes the atmospheric scattering model.

Here, I is the observed hazy image, J is the true scene radiance, A is the global atmospheric light, t is the medium transmission map, and x is the index pixel. In addition, J(x) can be reconstructed into I(x) by A and t, and t(x) refers to the portion of light that reaches the camera without being scattered. Moreover, t(x) is composed of the depth d(x) and the scattering coefficient of atmosphere β and is defined through Equation (2).

Because it is impossible to measure all variables of the atmospheric scattering model using current sensors, most studies on haze limit the number of variables to two.

Dehazing methods with single image inputs can be separated into methods using information in the image and learning-based approaches. The method using information in the image starts with a model for estimating the optical transmission [9], the dark channel prior (DCP) [10] based on the statistics of haze-free images, and the color attenuation prior (CAP) [11] for reconstructing the depth information. Many other studies are also ongoing. Learning-based methods using neural networks, which exhibit an excellent performance in various fields where images are input, generally include CNN- and GAN-based approaches. The CNN-based method aims to reconstruct the true scene radiance by estimating the global atmospheric light and the medium transmission map of the atmospheric scattering model [12]-[14]. These methods comprise stacking a large number of convolutional layers, and thus an estimate takes an excessive amount of time to achieve because the number of parameters is large.

Image dehazing is typically composed of a stack of convolutional layers, requiring a high-performance GPU. For this reason, it is difficult to enter all images into the dehazing network when constructing the dehazing system. Therefore, to reduce the load on the hardware, a haze detection network that detects the presence of haze and determines whether a dehazing network must be used is required. Studies on haze detection have been conducted in parallel with haze removal, and methods using image processing include the dark channel prior [10], which assumes colorful images, and a fast semi-inverse approach [17] that generalizes pixel-based operation. As neural networks continue to be applied to various fields, research on haze using CNN and GAN methods will increase.

In this paper, for a real-time automated monitoring system, we propose a selective dehazing system that determines the existence of haze and removes it from the image.

The proposed system is shown in Figure 1, and the haze detection network determines whether haze exists in the input image. If haze is present, it is input into the dehazing network and outputs the image with the haze removed. If haze does not exist, the input image is output. The detection network consists of a convolutional layer and a fully connected layer, and outputs the probability of haze to determine whether it is input to the dehazing network. The dehazing network is an encoder–decoder architecture that is an improved U-Net [18] and removes haze in the input image according to the judgment of the detection network.

Figure 1: 
Overall framework of proposed method

The haze detection network outputs the probability of haze being present in the input image and uses only the minimum layer to maintain the performance and reduce the number of computations in the dehazing system.

Figure 2 shows the architecture of the haze detection network, which is composed of four convolutional layers and two fully connected layers to determine the probability of haze presence throughout the image for feature extraction. To extract local features in the image and minimize the number of computations, the kernel size of the convolutional layer is set to (7, 7), the stride is set to (3, 3), and the padding is not set. As a result, the number of parameters in the flattening process is minimized. The proposed network is trained for cross-entropy loss used in binary classification, and outputs a 1 when haze exists and a 0 otherwise.

Figure 2: 
Architecture of Haze Detection Network

A dehazing network is a network that estimates haze-free images from images where haze exists in the haze detection network.

Figure 3 shows the architecture of the dehazing network and is an encoder–decoder structure that extracts strong features in an image by increasing the number of channels while reducing the size of the image. The encoder consists of four convolutional blocks consisting of two layers, i.e., a convolutional layer and a max pooling layer, and the decoder consists of three convolutional blocks consisting of a concatenate layer, an up-sampling layer, and two convolutional layers. The concatenate layer of the decoder is an input data mirroring process that compensates for pixel loss in the image boundary area that occurs during down-sampling. The proposed network was trained using the mean squared error between a haze-free image and output, and using a large number of convolutional layers, it has a large number of computations, demonstrating an excellent dehazing performance.

Figure 3: 
Architecture of dehazing network

In the O-haze dataset [19], which is the input of the experiment, haze was generated using two commercial haze generating machines to create an environment in which the haze is evenly distributed. The dataset consists of 45 different outdoor images consisting of hazy and haze-free images. The dataset consists of 45 different outdoor images, consisting of a hazy image and a haze-free image.

The proposed network is based on Python 3.7 and Tensorflow 2.3.1, and the GPU is an NVIDIA QUADRO RTX 8000. In this experiment, the Adam Optimizer with a learning rate of 0.001 was used, and the batch size was 32. The number of training data was 35; the number of validation data was 5; and the number of testing data was 5. The complete learning process was applied 100 times.

Table 1 shows the haze detection network evaluation conducted on a total of 10 haze images as the input data of the system, and the ground truth, which is a haze-free image. If the output is greater than 0.5, the image has haze, and at less than 0.5, the image is judged as a haze-free image, and the haze detection performance is 100%. However, when inputting the ground truth of O-haze 36, the probability of an image having haze was estimated at 10%. This is believed to be because the ratio of the sky area in the image is high, as is the intensity. In addition, to verify the haze detection network, two images captured on a day of sea fog occurrence near Korea Maritime University were input, and both were 0.999, confirming the existence of haze.