计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (24): 227-237.DOI: 10.3778/j.issn.1002-8331.2207-0214

• 图形图像处理 • 上一篇    下一篇

多分支修复网络的沙尘降质图像增强算法

丁元,邬开俊   

  1. 兰州交通大学 电子与信息工程学院,兰州 730070
  • 出版日期:2023-12-15 发布日期:2023-12-15

Sand Dust Degradation Images Enhancement Algorithm via Multi-Branch Restoration Network

DING Yuan, WU Kaijun   

  1. School of Electronics and Information Engineering, Lanzhou Jiaotong University, Lanzhou 730070, China
  • Online:2023-12-15 Published:2023-12-15

摘要: 目前基于深度学习的沙尘图像增强算法大多类似于图像去雾算法,但由于以下两个方面的原因效果并不理想且弱于一些传统图像处理算法:由于沙尘图像整体色彩偏黄,去雾算法会忽略沙尘图像的色彩恢复问题;由于缺乏大规模基准数据集,深度神经网络在有限的数据上学习从沙尘图像到清晰图像的映射是非常困难的。提出一种基于多分支修复网络的沙尘降质图像增强算法;此外,基于大气散射模型构建了一个新型沙尘图像数据集。算法将神经网络分为三个子网,包括迁移学习子网、色彩恢复子网和数据拟合子网,每个子网有其特殊的作用,沙尘图像分别经过三个子网处理,然后将三子网结果通过一个可学习的融合层映射为清晰图像。实验结果中定性比较表明该方法可以有效恢复沙尘图像细节,并较好恢复图像的视觉色彩,且该方法对比其他先进的方法可以产生更加符合人眼视觉体验的清晰图像;从定量比较中,在合成数据集上提出的算法相比于所对比的先进算法PSNR和SSIM指数分别提高了0.783和0.012,在真实图像数据集上提出的算法取得了最好的NIQE和PIQE指数。

关键词: 沙尘图像, 沙尘图像增强, 分支网络, 沙尘图像数据集, 颜色校正, 自适应归一化

Abstract: At present, most of the dust image enhancement algorithms based on deep learning are similar to image defogging algorithms, but the effect is not ideal and weaker than some traditional image processing algorithms due to the following two reasons. Firstly, because the overall color of the sand dust images is yellow, the dehazing algorithm will ignore the color recovery of the sand dust images. Secondly, due to the lack of large-scale benchmark data set, it is very difficult for deep neural network to learn the mapping from sand dust images to clear images on limited data. Therefore, this paper proposes a sand dust degradation images enhancement algorithm based on multi-branch repair network. In addition, a new sand dust images dataset is constructed based on atmospheric scattering model. The algorithm divides the neural network into three subnets, including transfer learning subnet, color recovery subnet and data fitting subnet. Each subnet has its special role. The sand dust images are processed by three subnets, and then the three subnet results are mapped to a clear images through a learning fusion layer. The qualitative comparison in the experimental results shows that this method can effectively restore the details of the dust image and better restore the visual color of the image, and this method can produce a clearer image that is more in line with the human visual experience than other advanced methods. From the quantitative comparison, the PSNR and SSIM indexes of the proposed algorithm on the synthetic dataset are 0.783 and 0.012 higher than those of the advanced algorithms, respectively. The proposed algorithm achieves the best NIQE and PIQE indexes on the real image dataset.

Key words: sand dust images, sand dust images enhancement, branch network, sand dust image datasets, color correction, adaptive normalization