基于生成对抗网络的图像去雾算法

In recent years, haze weather has frequently occurred. In the haze environment, the images obtained by the video acquisition equipment will be degraded due to the scattering effect of suspended particles in the atmosphere, and the video perception faces quality problems. In order to solve the above problems, we propose a dehazing method, which can effectively dehaze the hazy images obtained by the video capture device. Our approach adopts an end-to-end model based on generative adversarial networks to achieve image style transfer from hazy to dehaze images.For better dehazing effect, the model adds an attention mechanism on the basis of the generative adversarial network, and uses the proposed haze concentration formula to make the attention feature pay more attention to the fog part and achieve effective dehazing. At the same time, most of the hazy images in the existing dehazing datasets are synthesized, which is different from the real hazy environment, and the generalization is poor in practical applications. We use recurrent generative adversarial networks to construct pairs of real hazy datasets. The effect of the model trained on the new dataset is closer to the real dehazing scene and has better scene versatility.