Satellite Multi-local Component Detection And Low-light Image Enhancement Method

For satellite interaction missions such as autonomous docking,the key technology of these missions is satellite components detection,Deep-learning-based satellite component detection method has gradually become a research hotspot and has achieved great improvement in detection accuracy.In this paper,satellite components detection methods at home and abroad are investigated and analyzed,and the target detection,instance segmentation,and low-light image enhancement technologies are studied in depth.Finally,the deep-learning-based method is adopted to realize satellite components detection,segmentation,and low-light image enhancement.The main work of this paper is summarized as follows:1.In the field of satellite component detection and segmentation,because of a lack of public real satellite datasets,this paper proposes a multi-perspective and multi-position of the satellite appearance sampling method.Based on the simulation image,we build an information-rich satellite components dataset,which can be used for components detection and segmentation model raining and testing.Through the proposed multi-view and multi-orbit satellite appearance sampling method,the collected simulation images can more fully simulate the appearance of the satellite in the real scene,to obtain more abundant and effective samples,and provide more powerful support for the learning and performance testing of the model.2.Aimed at the problem of low precision of traditional satellite components detection methods,this paper proposed a detection method based on convolution neural networks.The residual connection structure,feature pyramid structure,and dense connection structure are fused and introduced into the Mask R-CNN for improvement,which improves the accuracy and recall of components detection.On this basis,to solve the problems of low detection/segmentation accuracy and poor generalization ability of deep-learning-based model caused by a small number of samples,this paper improves the target detector,attention mechanism,and loss functions on the basis of FCOS and CenterMask,and proposes a satellite multi-components segmentation method.Under the condition of a small number of samples,the autonomous and accurate target components detection and segmentation are realized.Experimental results show that the proposed component detection and segmentation methods have better performance than the original networks in terms of detection/segmentation accuracy and speed.3.To solve the problem of low image quality caused by low illumination imaging conditions in space,this paper proposes a low-light image enhancement method for space satellites based on generative adversarial networks.The designed U-Net structure with enhanced information transmission is introduced into the EnlightenGAN model,and the training of the model can be realized through unpaired samples.Experimental results show that the proposed method can not only better improve the brightness and contrast of low-light images but also better recover the image details.