基于改进循环生成式对抗网络的图像去雾方法

针对现有图像去雾方法存在的颜色失真、细节丢失以及去雾效果不自然等问题，提出一种改进的循环生成式对抗网络用于图像去雾。通过添加多尺度鉴别器作为判别器来改进原始网络结构，增强判别能力，引导网络产生更精细自然的无雾图像。同时重新设计损失函数，使用最小二乘代替交叉熵作为对抗损失，引入循环感知损失，结合原始循环一致性损失组成新的复合损失函数，提高图像颜色与细节恢复的质量。在D-HAZY和SOTS数据集上的实验结果表明：该方法能够生成较为自然的无雾图像，其主观效果和客观指标均优于对比方法，具有更好的去雾能力；与原始循环生成式对抗网络相比，峰值信噪比从19.052 dB提高至23.128 dB，结构相似性指数从0.787提高至0.867。与DehazeNet、AOD-Net与GCANet等主流去雾方法相比，峰值信噪比和结构相似性指数比排名第二的方法分别提升7.1%和4.3%。

Image Defogging Method Based on Improved Cycle-Consistent Adversarial Network

Aiming at the problems of color distortion, loss of detail, and unnatural defogging effects in the existing image defogging methods, an image defogging method based on improved Cycle-consistent Adversarial Network(CycleGAN) is proposed.This method improves the original network structure and enhances the discrimination ability by adding a multi-scale discriminator to guide the network to produce a finer and more natural fog-free image.Meanwhile, the loss function is redesigned, the least squares method is used to replace the cross-entropy as the counter loss, and the cyclic perception loss is introduced.Combined with the original cyclic consistency loss, a new composite loss function is constructed to improve the quality of the image color and detail restoration.Experimental results on D-HAZY and SOTS datasets show that the proposed method can generate more natural fog-free images, and its subjective effect and objective index are better than those of other methods and has better defogging ability.Compared with the original CycleGAN, the Peak Signal to Noise Ratio(PSNR) of the method is improved from 19.052 dB to 23.128 dB and the Structure Similarity Index(SSIM) is improved from 0.787 to 0.867.Compared with mainstream defogging methods such as DehazeNet, AOD-Net and GCANet, the PSNR and SSIM of this method are approximately 7.1% and 4.3% higher than those of the second-ranking algorithm, respectively.