摘要
雾降低了图像的清晰程度和细节信息,从而对后续视觉信息处理的影响是一个具有挑战性的问题。现有的图像去雾算法虽能去除雾度,但对处理户外浓雾场景的效果较差。为了突破现有去雾算法的局限性,结合大气散射物理模型,提出了一种端到端的多尺度并行融合去雾网络。采用多尺度卷积从整体到局部提取不同尺度的特征,并对这些特征进行多次并行融合;通过引入残差模块对细节特征进行深入学习,可以恢复出更多的图像细节。实验结果和数据分析表明,提出的方法在合成图像和真实图像上都能表现出良好的去雾性能,PSNR和SSIM指标平均同比提升3%。
Haze reduces the clarity and details of the image, and thus the impact on subsequent visual information processing is a challenging issue. Although the existing image dehazing algorithm can remove the haze, it is less effective for processing outdoor dense fog scenes. In order to break through the limitations of the existing defogging algorithm, this research combined with the atmospheric scattering physical model, proposed an end-to-end multi-scale parallel fusion defogging network. The network uses multi-scale convolution to extract features of different scales from the whole to the part, and fuse these features in parallel multiple times. In addition, by introducing a residual module to carry out in-depth learning of detailed features, more image details can be recovered. Experimental results and data analysis show that the proposed method can exhibit good defogging performance on both synthetic and real images, with PSNR and SSIM indicators increasing by an average of 3% year-on-year.
作者
张鑫
娄小平
黄子岩
张文玥
ZHANG Xin;LOU Xiaoping;HUANG Ziyan;ZHANG Wenyue(Beijing Key Laboratory of Optoelectronic Testing Technology,Beijing Information Science and Technology University,Beijing 100192,China)
出处
《光学技术》
CAS
CSCD
北大核心
2020年第6期707-711,共5页
Optical Technique
基金
国家自然科学基金(51475047)
北京市自然科学基金B类重点课题资助项目(16JC0011)。
关键词
图像去雾
卷积神经网络
残差学习
深度学习
image dehazing
convolutional neural network
residual learning
deep learning
