3D reconstruction of space targets can provide prior structural information for space services, which is a key technology for improving system autonomy. Conventional 3D reconstruction methods rely on handcrafted features to recover the 3D structure of objects by dense matching. Therefore, affected by the symmetrical structure and non-Lambert imaging of spatial targets, conventional 3D reconstruction methods often suffer from mismatching and insufficient matches of feature points, resulting in a low reconstruction accuracy. In recent years, with continuous developments in deep learning technology, convolution neural networks (CNNs) have been widely used in computer vision. Compared with the handcrafted features used by conventional 3D reconstruction methods, the deep features extracted by CNNs can introduce high-level semantics of images for more robust matching. Inspired by this, a 3D reconstruction method based on MVSNet for space targets is proposed. This algorithm organically applies CNNs with different structures to improve the accuracy and completeness of 3D reconstruction. We hope that our basic strategy and findings will be beneficial to the 3D reconstruction of space targets.

The space-target 3D-reconstruction algorithm model (Fig. 1) is described as follows. First, in view of the imaging characteristics of a space target, the influence of the geometric structure and material of the model on reconstruction results is analyzed, and a multi-view acquisition system for space targets based on the Blender platform is designed. Subsequently, deep visual image features are fully extracted via multi-scale convolution based on MVSNet. The coder and decoder are then used to gather and regularize the spatial context information for stereo matching, which effectively avoids the heavy dependence of conventional methods on the feature points in the reconstructions of low-textured, reflective, and repetitive texture regions. Finally, the residual network is used to solve the boundary smoothing problem caused by the multiple convolutions to further improve the reconstruction results. The model is tested on both the DTU dataset and our self-collected space-target dataset. Its performance is compared with those of VisualSFM, COLMAP, and SurfaceNet through both qualitative and quantitative evaluations. The running time of the proposed algorithm and other methods is also measured to verify the efficiency improvement. In addition, the influence of different numbers of matching views on the accuracy of the reconstructed model is studied to discuss the most appropriate settings.

The proposed method outperforms conventional 3D reconstruction methods in handling low-textured, specular, and reflective regions, which can completely restore typical structures, such as satellite cabins and roofs (Figs. 6 and 7). The mean accuracy, mean completeness, and overall errors of the proposed algorithm are 0.449, 0.379, and 0.414 mm, respectively. The proposed algorithm has the best accuracies among all four compared algorithms (Table 1). In particular, the accuracy of the proposed method is 20% higher than that of the advanced open source software COLMAP. The running time study shows that our method is faster with an average time cost of approximately 230 s for reconstructing one scan (Table 2). The running speed of the proposed method is 100 times faster than that of COLMAP and 160 times faster than that of SurfaceNet. In addition, the performance study on different numbers of matching views shows that more views result in a better performance (Table 3). However, the accuracy improvement is the greatest when three matching views are used for 3D reconstruction. Thus, the matching view number is set as 3 for model training. In general, the proposed model is optimal in terms of reconstruction accuracy and speed.

This study proposes a method for the 3D reconstruction of spatial targets based on the MVSNet deep learning network. First, deep visual image features are fully extracted by the multi-scale 2D convolution, and then spatial context information for stereo matching is fully gathered and regularized through the skip connection between the coding and decoding paths. Subsequently, the matching cost is converted into the depth value probability using the SoftMax function, and the expectation is calculated as the initial estimation value. Finally, the final depth estimation map is obtained by strengthening the edge semantic information through the residual network. Experimental results show that the accuracy of the proposed method is 20% higher than that of the advanced open source software COLMAP. Moreover, the running speed is 100 times faster than that of COLMAP and 160 times faster than that of SurfaceNet. In general, this model can effectively introduce the high-level semantics of images for more robust matching and has the low system running time, which can provide a technical reference for space operation automations and further promote the application of 3D reconstruction in this field.