结合强监督学习和生成对抗网络的图像去雾

Image dehazing combining strongly-supervised learning and generative adversarial network

针对现有去雾算法的复原图像易出现颜色失真与细节丢失问题,提出了一种基于改进循环生成对抗网络(cycle-consistent generative adversarial networks,CycleGAN)的端到端图像去雾方法,并无需依赖于大气散射模型的约束。网络生成器整体采用Encoder-Decoder架构,同时为有效学习有雾图像与清晰图像间的映射关系,在训练优化目标中结合图像自身属性构建了增强的高频损失与特征损失函数,实现对不同数据域的特征鉴别并进一步保证图像纹理结构。此外为约束复原图像与真实清晰图像颜色的一致性,提出了二阶段学习策略。首先通过非配对数据集对改进CycleGAN进行弱监督训练学习,然后于第二阶段利用部分成对数据集以强监督方式训练正向生成网络,在提高去雾网络稳定性的同时,使复原效果更接近于真实清晰图像风格。实验结果表明,所提去雾方法的峰值信噪比(peak signal to noise ratio,PSNR)和结构相似性(structural similarity,SSIM)指标值相比同类CycleGAN算法分别提升了12.43%与5.53%,并且同其他方法在视觉效果与量化指标的对比结果中也验证了其性能的有效性。

Aiming at the problems that the restored images of existing dehazing algorithms are prone to color distortion and details lost,an end-to-end image dehazing method based on cycle-consistent generative adversarial networks(CycleGAN)is proposed,which does not rely on the constriction of atmospheric scattering model.The whole network generator adopts the Encoder-Decoder framework,and in order to effectively learn the mapping relationship between hazy images and clear images,the enhanced high-frequency loss and feature loss functions are constructed in the training optimization objective by combining the image attributes,which realizes the feature identification of different data domains and further ensures the texture structure of images.In addition,a two-stage learning strategy is proposed to constrict the color consistency between restored images and their corresponding clear images.Firstly,the improved CycleGAN is learned by weakly-supervised training with unpaired datasets.Then in the second stage,the forward generator network is trained in a strongly-supervised manner with some paired datasets,which improves the stability of dehazing network while making the restoration effect closer to the real clear image style.The experimental results show that the peak signal to noise ratio(PSNR)and structural similarity(SSIM)of the proposed dehazing method are improved by 12.43%and 5.53%,respectively,compared with the similar CycleGAN algorithm,and the effectiveness of the proposed method is also verified by comparing the results of visual effect and quantitative index with other methods.