基于复合神经网络提升亚米级卫星影像质量

Improving Image Quality of Sub-Meter Satellite Based on Composite Neural Networks

低能见度情况下,大气粒子对太阳辐射的散射和吸收效应,降低了卫星影像像质和空间分辨率,传统图像处理方法和现在普遍应用的深度学习算法无法同时提升图像像质和空间分辨率。为了改变该现状,文章提出了基于网格去雾网络(GridDehazeNet)和真实超分辨率生成对抗网络(Real-ESRGAN)组合的复合神经网络。首先采用GridDehazeNet卷积神经网络架构提升卫星影像的清晰度和对比度,再利用Real-ESRGAN增强型超分辨率生成对抗网络以提升卫星影像空间分辨率;最后利用Worldview-3多光谱图像对不同算法进行了测试,并对比不同算法的测试效果。结果表明:该复合神经网络在改善图像像质和分辨率方面效果显著,其中清晰度提高了39.11倍,对比度提高了3倍,信息熵值提高了34%;且同时避免了传统算法所带来过度增强和噪声问题,对小目标物的识别和解译的准确率有显著提高。

In the case of low visibility,the scattering and absorption effect of atmospheric particles on solar radiation reduces the quality and spatial distribution of satellite images.Traditional image processing methods and deep learning algorithms commonly used today cannot simultaneously improve image quality and spatial resolution.In response to the current situation,the article proposes a composite neural network based on the combination of grid dehazing network(GridDehazeNet)and real-world enhanced super-resolution generative adversarial network(Real-ESRGAN)to improve image quality and spatial separation rate.In the research process of this paper,First,the GridDehazeNet structure neural network architecture is used to improve the clarity and contrast of satellite images,and then the Real-ESRGAN enhanced super-resolution adversarial network is used to improve the spatial resolution of satellite images;finally,different algorithms are tested using Worldview-3 multi-spectral images,and the test results of different algorithms are compared.The results show that the composite neural network has a significant effect on improving the image quality and resolution,among which the clarity is increased by 39.11 times,the contrast is increased by 3 times,and the entropy value is increased by 34%.Intensity and noise problems,the accuracy of recognition and interpretation of small targets has been significantly improved.