Adversarial dehazing network based on multi-scale dilated convolution

Images taken in hazy days suffer the problems of color distortion and blurred image details, which lead to a negative impact on the image quality captured by imaging equipment. Aiming at solving the degradation problem of image collection in hazy weather, a dehazing network based on multi-scale dilated convolution was proposed. The generator of the dehazing network was composed of convolution modules with different dilated rates, and combined with multi-scale strategies to increase the receptive field and enhanced the dehazing effect; the discriminator used multiple convolution modules to distinguish the generated dehazed images from real haze-free images; by calculating the perception distance between the dehazed images and the real haze-free images, the texture structure of the images was optimized and the noise signal was reduced. The experimental results show that the peak signal-to-noise ratio obtained by the proposed algorithm on the public data set is 22.410 dB, the structural similarity value is 0.844, and the color difference value is 10.545. Quantitative and qualitative evaluations show that the dehazing network designed with dilated convolution and perceptual loss can effectively restore the color information and texture structure of the images.