近年来，人类太空活动日益频繁，开发利用太空环境的需求不断增多，面向空间在轨服务的自主操作控制技术得到了极大关注，视觉感知作为其中的关键环节，具有很强的研究价值。在轨服务操作的对象大多为非合作目标，随着任务智能化要求的不断提升，传统的感知方法往往无法满足普适性要求。本文考虑空间在轨服务的实际情况，将视觉感知划分为特征检测和位姿估计两个子任务，分别基于深度学习方法开展技术研究，主要包括了以下工作：（1）针对特征检测问题，对主流的目标检测框架原理进行分析，考虑到后续的操作过程中可能需要对特征进行边缘提取和三维重建，因此选择将Mask R-CNN方法应用于空间非合作目标特征检测，在检测的同时完成特征的像素级分割。实验结果表明，Mask R-CNN方法可以有效检测出特征并对其完成分割。针对Mask R-CNN方法速度偏慢的问题，引入位置敏感得分图对算法进行优化改进，提升了检测速度。实验结果表明，与传统方法相比，改进的Mask R-CNN方法可缩短20%的检测时间，能更好地满足空间任务实时性要求。（2）针对位姿估计问题，对比分析了有监督学习和无监督学习两种方法：有监督学习将位姿估计简化为回归问题，从图像直接映射出目标的绝对位姿，在不同场景中普适性较差；无监督方法以投影图像与目标图像的重构误差作为监督信息，隐式地学习图像间的相互关系和内在联系，有较强的普适性，但只能输出图像间的相对位姿。综合以上两种方法，本文提出一种半监督学习方法，在利用位姿数据进行监督学习的同时利用图像间的重构误差学习图像间的相互关系和内在联系，增强算法普适性。实验结果表明，该方法的精度和速度均能较好地满足空间任务要求。（3）针对深度神经网络需要大规模数据集进行训练的特点，本文以迁移学习为理论支撑，基于虚拟环境搭建了空间目标感知任务数据集构建系统，构造了适用于本文研究任务的空间目标特征检测与位姿估计数据集。实验结果表明，网络可以在生成的数据集上完成训练，并且具备迁移到实际任务的能力。本文提出的数据集构建系统可以作为一种数据集构建的通用平台，迁移应用到其他场景。

In recent years, human space activities have become more and more frequent, and the demand of developing and utilizing space environment has been increasing. The autonomous operation control technology of On-Orbit Servicing has received great attention. As the key technology, visual perception has strong research value. The objects of on-orbit service operations are mostly non-cooperative target. With the continuous improvement of task intelligence requirements, traditional perception methods often fail to meet the universality requirements. In this paper, considering the actual situation of space on-orbit service, visual perception is divided into two sub tasks: feature detection and pose estimation, which are based on deep learning to carry out technical research, mainly including the following work:(1) For the problem of feature detection, this paper analyzes the mainstream target detection framework principles. Considering the edge extraction and 3D reconstruction of features may be needed in the following operations, this paper chooses to apply Mask R-CNN method to the feature detection of spatial non-cooperative target, which can complete the features’ pixel-level segmentation at the same time of detection. Experimental results show that Mask R-CNN method can effectively detect and segment features. Aiming at the problem of slow speed of R-CNN method, the position-sensitive score map is introduced to optimize the algorithm and improve the detection speed. The experimental results show that compared with the traditional method, the improved Mask R-CNN method can reduce the detection time by 20%, which can better meet the real-time requirements of space missions.(2) For the problem of pose estimation, this paper compares and analyzes supervised and unsupervised learning method: the former simplifies pose estimation into the regression problem. Mapping the absolute pose of the target from the image, the universality is poor in different scenes; The latter uses the reconstruction error between the projection image and the target image as the supervisory information and implicitly learns the interrelationship and internal relationship between the images, which has strong universality, but the output is relative pose between images. Combining the above two method, this paper proposes a semi-supervised learning method, which uses pose data for supervised learning and the reconstruction error between images to learn the relationship and internal relationship between images, so as to enhance the universality of the algorithm. Experimental results show that the accuracy and speed of this method can meet the requirements of space mission.(3) In view of the characteristics of deep neural network that needs large-scale dataset for training, this paper uses transfer learning as the theoretical support, proposes a dataset construction system of spatial target perception tasks based on virtual environments, constructs dataset of spatial target feature detection and pose estimation suitable for this research task. The experimental results show that network can complete the training on the generated dataset and has the ability to migrate to the actual task. The dataset building system proposed in this paper can be used as a general platform for data set building, which can be migrated and applied to other scenarios.