基于条件生成对抗网络的图像去雾方法  

Image dehazing method based on conditional generative adversarial network

作　　者：张婷[1] 赵杏 陈文欣 ZHANG Ting;ZHAO Xing;CHEN Wenxin(School of Automation,Beijing Institute of Technology,Beijing 100081,China)

机构地区：[1]北京理工大学自动化学院,北京100081

出　　处：《计算机应用》2021年第S02期248-253,共6页journal of Computer Applications

摘　　要：雾霾天气下拍摄收集的图像和视频能见度低,图像质量大幅下降,影响目标检测等高级计算机视觉任务检测精度。目前暗通道、DehazeNet、AOD-Net等去雾方法存在有雾残留、色差较大等问题。为解决以上问题,提出基于条件生成对抗网络(CGAN)的图像去雾方法——CGAN-dense。其生成器设计密集块结构以提高特征利用率,改善去雾细节保持效果;判别器使用34×34的PatchGAN进行分块判定,提高图像分辨率;损失函数增加内容损失,减少像素级别损失。在合成有雾数据集Hazy Dataset中,所提方法的峰值信噪比(PSNR)达到35.45 dB以上,结构相似度(SSIM)达到0.970 2以上;在Reside数据集中,PSNR达到22.25 dB以上,SSIM达到0.908 8以上,去雾后目标检测率相比去雾前提高6.96个百分点。实验结果表明,所提方法能够减少色差较大和有雾残留问题,提高了图像去雾的色彩保真性和细节信息保持度。Images and videos have low visibility in hazy days and the quality reduces greatly simultaneously,which affects the processing accuracy of advanced computer vision tasks such as object detection. Current dehazing methods,Dark Channel Prior(DCP),DehazeNet,AOD-Net,have haze residue or large color distortion. In order to solve above problems,a dehazing method CGAN-dense based on Conditional Generative Adversarial Network(CGAN)was proposed. To increase feature utilization rate and detail preservation,the dense block was designed in generator network. 34×34 PatchGAN was adopted by discriminator network,which improves the resolution. At the same time,content loss was added to decrease the pixel loss. On synthesized Hazy Dataset,dehazing Peak Signal-to-Noise Ratio(PSNR)of proposed method reaches more than 35. 45 dB,and Structural SIMilarity(SSIM)reaches more than 0. 970 2. On Reside Dataset,PSNR reaches more than22. 25 dB,and SSIM reaches more than 0. 908 8. Object detection rate is improved 6. 96 percentage points after dehazing.Experimental results show the proposed method can reduce color distortion and haze residue,and improve color and detail information of image.

关 键 词：图像去雾 条件生成对抗网络 目标检测 密集块 内容损失 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]