Image Hazing Algorithm Based on Generative Adversarial Networks

Abstract

Haze is an important factor in photography with a special aesthetic, emotional, or compositional meaning, an image hazing method is proposed based on generative adversarial network. The proposed network consists of two parts: the generator has a symmetric encoder and decoder structure with skip connection, which is used to generate hazy images; and the discriminator is a global fully convolutional network, which is used to identify the reality of the generated hazy images. The haze-free image is used as the input of the generative network to obtain a corresponding hazy image. Then the discriminative network judges the similarity between the original haze-free image and the corresponding hazy image to optimize the network parameters, which means that all the parameters of the entire network can be automatically learned through training. The loss function of the network is the combination of the GAN(Generative Adversarial Networks) loss and the REG(Regression) loss with a coefficient λ. We design a REG loss which includes feature loss loss FL and image loss loss IL. By calculating the loss, a model mapping the relationship between the two corresponding pixels from the original image and the synthetic hazy image respectively is well trained. In the last part of the paper, we have shown some impressing synthetic hazy images obtained by our model as well as analyzing the effects on images with different scenes. Moreover, we have contrasted the experimental results of our algorithm with that of other classical ones. Experiments demonstrate that the proposed algorithm has a better performance than other state-of-the-art methods on both virtual-scene and real-world images qualitatively and quantitatively. INDEX TERMS Generative adversarial networks, image hazing, regression loss, mapping from pixels to pixels.

An Approach to Deepfake Video Detection Based on ACO-PSO Features and Deep Learning

“…Sometimes, deepfake video frames are unclear, and to enhance their quality, we need to use the image haze removal method [25].…”

Research on algorithm of students' classroom behavior recognition based on two-stream convolutional neural network

“…The two-stream network model is shown in Figure 3. First, python, Opencv and other tools were used to extract the data set from RGB and optical flow graphs: The size of the input image is 224×224 [6]. When training, the tensor size of the RGB image input by the spatial network can be expressed as (3×N,224,224).…”

Rapid Fog-Removal Strategies for Traffic Environments

“…In order to further verify the effectiveness and advancement of Optimization Strategy 3 (the fusion of dark channel prior and Optimization Strategy 3), the performance of our optimized defog algorithm was compared with that of DCP [3], DCPDN [41], AOD-NET [18], CAP [42], EN-DCP [43] defog algorithms. We obtain the similarity index measure (SSIM) [44], and the peak signal-to-noise ratio (PSNR) [45], in order to calculate the average gradient of a foggy picture. We first convert the color image to a grayscale image, then use Sobel [46] to calculate the gradient on the X and Y axes of the image, then use the Euclidian distance [47] formula to calculate the gradient amplitude and, finally, calculate the average gradient amplitude, to obtain the average gradient.…”