光学学报, 2019, 39(12): 1228001, 网络出版: 2019-12-01

综合深度卷积神经网络的摆扫影像反演恢复算法

Inversion Restoring Algorithm for Whiskbroom Scanning Images Synthesized with Deep Convolutional Neural Network
作者单位

1中国科学院长春光学精密机械与物理研究所空间新技术研究部, 吉林 长春 130033

2中国科学院大学材料科学与光电技术学院, 北京 100049

摘要
为解决卫星摆扫获得的遥感影像存在畸变和像质退化的问题,提出一种分辨率反演与深度卷积网络相结合的几何校正与图像增强方法。摆扫过程中,空间相机的摆扫角和单位视场角恒定,探测器像面的像素与相机光轴指向的地面景物一一对应,根据这两点,像面上畸变的景物可以被精确地反演恢复。其次,采用真实的遥感影像作为样本,训练了针对遥感影像的深度卷积网络框架,解决了反演恢复过程中的影像模糊问题,增强了校正后影像的视觉效果。实验中,畸变校正后的影像很大程度地恢复了地面景物的原有几何特性。采用无参考图像质量评价指标(NR-IQA )对本文提出的网络结构、在传统图像集上训练的网络结构及插值方法进行对比评价,结果表明所提网络结构的图像增强结果更佳,能较为有效地恢复遥感摆扫影像。
Abstract
To overcome the limitation of distortion and quality deterioration in whiskbroom scanning images, we propose a geometric correction and image enhancement method that combines the resolution inversion with deep convolutional neural network (DCNN) architecture. During the whiskbroom scanning process, the total whiskbroom scanning angle and unit field of view angle of a space camera are invariable, and each pixel of the detector on the image plane corresponds to the ground scene pointed by the camera boresight. Suitably, these help in restoring compressed pixels accurately. Furthermore, we adopt real-scene remote sensing panchromatic images as the sample to train the DCNN for remote sensing panchromatic images. Then, image blurring during the process of inversion is solved, and the visual effect of the corrected image is enhanced. In our experiment, the distortion corrected imagery restores the geometric characteristics of the ground scene to a large extent. The no-reference image quality evaluation indicators are used to evaluate our proposed network architecture, network trained on generic image set and interpolation method. The experimental result indicates that our proposed network realizes the best performance of image enhancement among the three methods with a great restoration effect of the whiskbroom scanning images.
00 11