Single Image Dehazing Using Generative Adversarial Networks Based On Multi-scale Fusion Enhancement

In a hazy environment,due to the scattering of solar radiation caused by particles suspended in the air,the images collected by the image acquisition system will be significantly degraded,such as low color saturation,unclear imaging,blurred texture details,etc.As a result,accurate information cannot be obtained from the image.At present,the single image dehazing algorithms are mainly divided into enhancement-based algorithms and restoration-based algorithms.The enhancement-based dehazing algorithms do not consider the root cause of the image quality degradation caused by the interference of haze weather,resulting in an unnatural image after dehazing.The restoration-based dehazing algorithm is divided into the algorithm based on prior knowledge and the algorithm based on deep learning.The former is not suitable for the changeable hazy environment and has great limitations,and the latter is divided into relying on the atmospheric scattering model.The algorithm relying on the atmospheric scattering model is prone to errors in parameter estimation,and the end-to-end dehazing algorithm is better based on the dehazing algorithm based on the generative adversarial network,which uses the game idea to make the generator and the discriminator fight against each other.It has strong generalization performance.This paper mainly studies the image dehazing algorithm based on generative adversarial network.The main work is as follows:(1)Due to the blur of images collected in hazy days,the restored dehazed images lose detailed texture features.Therefore,an enhanced multi-scale generative adversarial network model based on pix2pixHD is proposed.The EPM-GAN model is an end-to-end network model without relying on atmospheric scattering models and prior knowledge.It generates coarse-to-fine dehazed images at both global and local scales through a multi-scale discriminator network supervised by a multi-level generator network,and embeds a multi-scale enhancer in the multi-level generator to make the fusion of dehazed images more efficient multi-detail texture features.Using the RESIDE dataset to train the EPM-GAN model,and conduct extensive testing experiments on the SOTS test set and real-world dataset,the PSNR value is improved by 1.73dB and the VI is improved by 0.0295.The experimental results verify the effectiveness of the EPM-GAN model in dehazing and detail feature restoration.(2)Through the analysis of several mainstream dehazing algorithms,the GCA-Net model has the problems that the dehazing image is too dark and the detail texture is blurred.Based on GCA-Net,a multi-scale generative adversarial network single image dehazing algorithm based on frequency fusion is proposed.MFF-GAN uses GCA-Net as the base network and embeds a discriminator network with a frequency prior.The discriminator network supervises the generator network in both high-frequency and low-frequency components to generate more detailed textured and natural dehazing images,where high and low frequencies are separated using a rolling guided filter.This paper uses the RESIDE data set to train the MFF-GAN model,and conducts extensive experiments on the SOTS test set.Whether it is based on qualitative analysis from different angles or quantitative analysis based on four evaluation indicators(Compared with GCA-Net,PSNR,SSIM,RI,VI are improved by 0.21dB,0.0438,0.0086,0.0229 respectively),the MFF-GAN model has achieved satisfactory results.