An improved lightweight model based on Mask R-CNN for satellite component recognition

Satellite component recognition has always been a hot topic in the field of orbital services. However, it is very challenging to segment the components such as satellite body, solar panel, and antenna in pixel-level accurately due to the poor illumination condition and the scarce image for spaceborne observation. Based on the Mask R-CNN, this paper proposes a lightweight instance segmentation model for satellite component segmentation and recognition. It improves residual module by using deep separable convolution, replacing nonlinear activation function with linear one after deep separable convolution and deleting the dimensionality reduction convolution layer in residual module. Also, the training datasets consist of the synthetic images generated by the 3D max software and the C-DCGAN based image generation method through several known satellite CAD models. The simulation experiments are carried out and the results show that the proposed method can effectively recognize the typical satellite components and achieve better performance than the compared model in aspects of accuracy, model parameters, and model size.