Journal of Computer Applications ›› 2020, Vol. 40 ›› Issue (3): 865-871.DOI: 10.11772/j.issn.1001-9081.2019071219

• Virtual reality and multimedia computing • Previous Articles     Next Articles

MP-CGAN: night single image dehazing algorithm based on Msmall-Patch training

WANG Yunfei, WANG Yuanyu   

  1. College of Information and Computer Science, Taiyuan University of Technology, Jinzhong Shanxi 030600, China
  • Received:2019-07-15 Revised:2019-09-04 Online:2020-03-10 Published:2019-09-19
  • Supported by:
    This work is partially supported by the Natural Science Foundation of Shanxi Province (201801D121142).

基于Msmall-Patch训练的夜晚单幅图像去雾算法——MP-CGAN

王云飞, 王园宇   

  1. 太原理工大学 信息与计算机学院, 山西 晋中 030600
  • 通讯作者: 王园宇
  • 作者简介:王云飞(1994-),男,山西晋中人,硕士研究生,主要研究方向:计算机视觉、计算智能;王园宇(1973-),男,山西汾阳人,副教授,博士,主要研究方向:计算机视觉、机器人、计算智能、虚拟现实。
  • 基金资助:
    山西省自然科学基金资助项目(201801D121142)。

Abstract: Aiming at the problems of color distortion and noise in night image dehazing based on Dark Channel Prior (DCP) and atmospheric scattering model method, a Conditional Generated Adversarial Network (CGAN) dehazing algorithm based on Msmall-Patch training (MP-CGAN) was proposed. Firstly, UNet and Densely connected convolutional Network (DenseNet) were combined into a UDNet (U Densely connected convolutional Network) as the generator network structure. Secondly, Msmall-Patch training was performed on the generator and discriminator networks, that was, multiple small penalty regions were extracted by using the Min-Pool or Max-Pool method for the final Patch of the discriminator. These regions were degraded or easily misjudged. And, severe penalty loss was proposed for these regions, that was, multiple maximum loss values in the discriminator output were selected as the loss. Finally, a new composite loss function was proposed by combining the severe loss function, the perceptual loss and the adversarial perceptual loss. On the test set, compared with the Haze Density Prediction Network algorithm (HDP-Net), the proposed algorithm has the PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural SIMilarity index) increased by 59% and 37% respectively; compared with the super-pixel algorithm, the proposed algorithm has the PSNR and SSIM increased by 59% and 48% respectively. The experimental results show that the proposed algorithm can reduce the noise artifacts generated during the CGAN training process, and improve the night image dehazing quality.

Key words: Dark Channel Prior (DCP), night image dehazing, Conditional Generative Adversarial Network (CGAN), perceptual loss, adversarial perceptual loss

摘要: 针对基于暗通道先验(DCP)与大气散射模型方法实现夜晚图像去雾出现颜色失真及噪声等问题,提出一种基于Msmall-Patch训练的条件生成对抗网络(CGAN)去雾算法MP-CGAN。首先,将UNet与密集神经网络(DenseNet)网络结合成UDNet网络作为生成器网络结构;其次,对生成器与鉴别器网络使用Msmall-Patch训练,即通过对鉴别器最后Patch部分采取Min-Pool或Max-Pool方式提取多个小惩罚区域,这些区域对应退化严重或容易被误判的区域,与之对应提出重度惩罚损失,即在鉴别器输出中选取数个最大损失值作为损失;最后,将重度惩罚损失、感知损失与对抗感知损失组合成新的复合损失函数。在测试集上,与雾密度图预测算法(HDP-Net)相比,所提算法的峰值信噪比(PSNR)与结构相似性(SSIM)值分别提升了59%与37%;与超像素算法比,PSNR与SSIM值分别提升了59%与48%。实验结果表明,所提算法能够减少CGAN训练过程产生的噪声伪影,提高了夜晚图像去雾质量。

关键词: 暗通道先验, 夜间图像去雾, 条件生成对抗网络, 感知损失, 对抗感知损失

CLC Number: