Optical sparse aperture (OSA) systems show great potential for the next generation astronomical telescope system due to its excellent high resolution with low volume and weight. However, the sparse arrangement causes its mid-frequency modulation transfer function to be lower compared with a single fully-filled aperture system, which further leads to blurred images and reduced contrast. Therefore, image restoration becomes an indispensable part for OSA systems. In this paper, a dual target network (DTN) is proposed for the image restoration of OSA systems. The noise in a raw image is estimated with interpolation and difference calculation. A block matching 3D filter is used as a denoiser. A denoised image is regarded as a degraded image which cannot be accurately modeled. To cope with the restoration problem, a dual target (negative structural similarity and the sum of fidelity and regularization term) network is trained. A function determined by the filling factor and the aperture distribution is trained as a correction term of the network. The trained network is used to deconvolve the denoised image. Simulation and experiment results show that the proposed method has good peak signal-to-noise ratio and structure similarity. For a Golay-6 system with a filling factor of 0.3245, when the signal-to-noise ratio is 30 dB, the DTN method increases the average peak signal to noise ratio from 22.6 dB to 31.7 dB and improves the average structural similarity from 0.77 to 0.90.

稀疏光学孔径（OSA）系统由于其出色的高分辨率，小体积和轻巧的特性，在下一代天文望远镜系统中显示出巨大潜力。然而，与单个完全填充的光圈系统相比，稀疏的排列导致其中频调制传递函数降低，这进一步导致图像模糊和对比度降低。因此，图像恢复成为OSA系统必不可少的部分。本文提出了一种用于OSA系统图像恢复的双目标网络（DTN）。原始图像中的噪声通过插值和差值计算进行估算。块匹配3D滤波器用作去噪器。去噪图像被认为是无法准确建模的降级图像。为了解决恢复问题，训练了双重目标（负结构相似性以及保真度和正则项之和）网络。由填充因子和孔径分布确定的函数被训练为网络的校正项。经过训练的网络用于对去噪图像进行去卷积。仿真和实验结果表明，该方法具有良好的峰值信噪比和结构相似性。对于填充因子为0.3245的Golay-6系统，当信噪比为30 dB时，DTN方法将平均峰值信噪比从22.6 dB提高到31.7 dB，并将平均结构相似度从0.77提高至0.90。由填充因子和孔径分布确定的函数被训练为网络的校正项。经过训练的网络用于对去噪图像进行去卷积。仿真和实验结果表明，该方法具有良好的峰值信噪比和结构相似性。对于填充因子为0.3245的Golay-6系统，当信噪比为30 dB时，DTN方法将平均峰值信噪比从22.6 dB提高到31.7 dB，并将平均结构相似度从0.77提高至0.90。由填充因子和孔径分布确定的函数被训练为网络的校正项。经过训练的网络用于对去噪图像进行去卷积。仿真和实验结果表明，该方法具有良好的峰值信噪比和结构相似性。对于填充因子为0.3245的Golay-6系统，当信噪比为30 dB时，DTN方法将平均峰值信噪比从22.6 dB提高到31.7 dB，并将平均结构相似度从0.77提高至0.90。