使用深度神经网络检测Cassini ISS图像中圆盘状星体轮廓

Cassini空间探测器光学成像系统（ISS）拍摄的图像中，很多卫星呈现为面元，其轮廓检测是天体测量的重要工作.使用神经网络方法进行ISS图像中面元轮廓检测.每个ISS图像的像素分为轮廓边缘和非轮廓两类.使用神经网络框架TensorFlow，输入每个像素的9个特征，输出每个像素的分类.利用约3.6万个像素训练该网络，通过380幅ISS图像进行测试.与人工标记结果相比，轮廓像素检测的平均精确率为78.26%，平均召回率为73.32%.以检出轮廓像素作为输入，通过椭圆拟合得到面元的轮廓，所得轮廓与面元真实轮廓吻合良好.研究结果表明该方案能够有效检测出面元轮廓，进而给出假图像星的排除范围. 

Contour Detection of Disk Resolved Objects in Cassini ISS Image Using Deep Neural Network

In the astrometry of CCD image,it is important to match image stars with catalogue stars to correct the camera’s pointing.The onboard Imaging Science Subsystem(ISS)in Cassini orbiter has taken a large number of images of targets which are disk resolved.In the astrometry of these images,the false image stars are often detected in the disk.It disturbs the pointing correction of camera and decline the precision of the astrometry.Therefore,it is helpful to find the contours of the disk to remove the false image stars.One method based on deep learning is proposed to detect the contour of disk resolved object in ISS images.A convolutional neural network was set up by the framework TensorFlow,in which the input is the nine features of each pixel,and the output is the classification of each pixel:contour pixel or non-contour pixel.The neural network is trained by about 36000 pixels,and then it is used to detect the contotr pixels in 380 ISS images.Compared with the contour pixels labeled by hand,the contour pixels detected by neural network have the precision of 78.26%and the recall ratio of 73.32%.It proved that the proposed method is available to find the contour of disk resolved target in ISS images.