Abstract:

To solve the problem that existing dehazing models are prone to over fitting after training with synthetic hazy image data sets, an image self-enhancement dehazing algorithm is proposed in combination with generative adversarial network. The depth information of an image is estimated while combining two Generative Adversarial Networks. The first GAN uses a clear image to learn the process of image hazing, and then adopts the hazed image generated by it as the input of the second GAN to guide the second GAN to correct dehazing. In order to reduce the difference before and after image processing, the consistency loss function is applied to optimize two network parts. In addition, the scene depth estimation module is added to the image hazing part, and the scattering factor is randomly sampled to achieve the image self enhancement function, which can more realistically simulate fog of different concentrations in real world. The algorithm does not need to use the paired information of synthetic hazy image dataset to further avoid over fitting. Experimental results show that the proposed algorithm can achieve better dehazing effect, which has good performance in subjective visual quality and objective evaluation indicators, superior to similar algorithms.

Key words: image processing, machine vision, generative adversarial network, optical model, image dehaze