Research On Intelligent Detection And Recognition Algorithm Of Space Target Based On Deep Learning

With the continuous innovation of satellite technology,some countries have proposed new goals for creating an air-ground integrated space situational awareness system.With the explosion of artificial intelligence technology,satellite applications using visible light images as carriers have ushered in new challenges and opportunities.Using deep learning to achieve intelligent space target detection and recognition has become a new research hotspot.However,there are currently several difficulties that need to be resolved.Firstly,the price of acquiring visible light satellite images is still very huge.The existing public datasets about satellite are all obtained by software simulation.The satellite angle and attitude in the image are relatively single,and its number and diversity are difficult to meet the training requirements of neural networks.Secondly,because satellite images in real scenes have the characteristics of wide format,strong noise interference,and some target pixels are discontinuous,a new method of space target detection is needed.Thirdly,because the external shape of the satellite in the space target image is relatively similar,the accuracy of target recognition directly on the entire image needs to be improved.Strengthening the use of the semantic information of its main components is the key to improving the recognition accuracy.In view of the above problems,this thesis has studied the generation,detection and recognition of space target images based on deep learning technology.First,this thesis implements a space target image generation algorithm C2 GAN based on image key points.The algorithm is based on the joint development of generative adversarial network,exploring key points and image data in an interactive manner.It is characterized by implicitly constraining each other through each cycle,which brings additional supervision across cycles,thereby promoting deep optimization of the entire network.Experiments show that the algorithm can generate multi-angle and multi-pose data samples that meet the training requirements based on key points of satellite.Second,this thesis proposes a space target detection algorithm based on threshold embedding.The algorithm consists of positioning and classification.The positioning algorithm is based on the minimum bounding rectangle algorithm,which adds a minimum fusion threshold.The algorithm will determine whether to merge them by calculating the intersection ratio between the rectangular boxes when multiple positioning boxes appear in an image.And then the algorithm improves the positioning accuracy by comparing the intersection and the given threshold.An end-to-end space target detection algorithm is formed by combining the positioning algorithm and the binary classification network for distinguishing satellites and space debris.Next,this thesis proposes a multi-granularity recognition algorithm for space targets fused with semantic information of components.The algorithm first performs pixel-level component segmentation on the space target,and then integrates the semantic information of the main components and the overall semantic information.By increasing the feature weight of the decision-making ratio,a higher level of satellite expression is obtained.Experiments show that the algorithm proposed in this thesis obtains higher recognition accuracy than the overall semantic information method.Finally,this thesis implements the coordinated space target detection and recognition system of each module according to the actual application requirements.Experiments show that the system can be used for rapid detection of space targets and accurate model recognition,which lays a technical foundation for the construction of intelligent space situation awareness system.