Research On Spatial Targets Recognition Neural Network Algorithm Based On Infrared Radiation Intensity Series

Target recognition based on infrared imaging is a crucial technique in missile defense systems.Infrared image of space targets at long-range distance are point targets,and detectors can only obtain the target radiation intensity sequences.How to extract feature information such as shape and attitude motion of the targets and carry out point targets recognition effectively are difficult problems of the research.Under the guidance of intelligent warfare requirements,this thesis studies the space target recognition neural network algorithms based on the infrared radiation intensity time series of space point targets.The major contributions of the thesis are as follows.Firstly,we constructe the space target infrared radiation intensity model and realize shape and attitude inversion.The characteristics of target surface temperature and attitude motion are modeled and analyzed,and the target projection area model and infrared radiation intensity sequence model are established.The factors affecting the target radiation intensity sequence are simulated and analyzed based on the model,and the radiation intensity sequences of four typical targets are generated as the data basis for subsequent recognition research.When the shape is uniquely determined,the shape parameters of the convex object are estimated by the projection area sequence,and the target shape inversion is realized.Furthermore,the joint estimation of the shape and attitude parameters of the target is carried out.For the non-convex objective function of the micro-motion parameter,the Chaotic Particle Swarm Optimization algorithm is adopted to avoid falling into local minimum values,and the target shape and attitude parameters are effectively estimated,thus realizing the inversion of the target shape and attitude.Secondly,according to requirements of space target recognition in long-range distance,we propose a classification algorithm of space targets based on independent random recurrent neural networks.On one hand,the independent random recurrent neural networks model adjusts the hidden layer structure.This measure solves the problem of gradient disappearance and explosion,enhances the ability to process long sequences,and is beneficial to learn long-term periodic features of sequence samples for classification.On the other hand,mapping historical information to the input space in a random weighted manner improves the memory ability and promotion performance of the model,and more importantly,enhances the classification ability of the model.The algorithm is tested on the UCR common dataset and the simulation dataset of the space targets.The experiment results show that the independent random recurrent neural networks model enhances the classification performance and the robustness to noise significantly compared with the basic recurrent neural networks.Finally,aiming at the condition of limited labeled samples in the problem of infrared radiation intensity time series classification of space target,we propose a space targets recognition algorithm based on semi-supervised auxiliary classification generative adversarial networks.The algorithm solves the problem of mode collapse during training effectively and improves the performance of the discriminator and generator.In view of the instability and non-convergence of the model in the training process,two improved measures of feature matching and virtual batch normalization are proposed.Experiments are conducted to test the classification performance and generation performance of the semi-supervised auxiliary classification generative adversarial network model.It is verified that the model can classify the spatial target infrared radiation intensity sequence effectively and has high robustness to noise.The generated samples have a high degree of similarity to the real samples.