Algorithm Research On Multi-Imformation Feature Fusion

The technology of multi-pattern composite guidance has become the main research direction of precision guidance weapon.Millimeter wave radar seeker has a long detection distance and can work all weather,but it is easy to be affected by electronic interference and electronic deception.Although infrared thermal imaging seeker has high detection accuracy and strong anti-jamming ability,its detection distance is very short.Millimeter wave radar / infrared thermal imaging dual-mode composite guidance system can make up for the shortcomings of single-mode guidance by using the advantages of each single-mode detection.Feature level information fusion technology can fuse multi-source feature information provided by radar and infrared detector,filter out useless features and redundant features,reduce feature space dimension,solve the problem of data heterogeneity to a certain extent,so as to improve the accuracy and efficiency of information fusion system in target classification task.So far,information fusion technology has achieved remarkable results at home and abroad,but there are still many problems to be solved for feature level information fusion technology,and it is urgent to carry out in-depth theoretical research to provide theoretical basis and technical support for the engineering realization of multimode composite seeker.On the basis of reference and summary of the existing algorithm results,combined with the particularity of missile platform,the radar / infrared multi-source information feature level fusion algorithm is studied more systematically and in-depth.The specific research contents are as follows:(1)This paper studies the training sample data of non-cooperative target from all angles.For the millimeter wave radar detector,the high resolution range profile of the target at all angles is simulated.By using the simulation method of High resolution range profile of radar noncooperative target based on the hybrid model,through building the fine scattering model of the target,and compared with the measured data,the results show that the scattering point position corresponds one by one,and the amplitude relationship between the strong scattering point and the weak scattering point is basically the same.For the infrared thermal imaging detector,the infrared image of the target corresponding to the radar angle is simulated.By using the infrared image simulation method based on thermal analysis,according to the three-dimensional model of the target,by solving the thermal equation and simulating the detector effect,it is found that the simulation and the measured infrared image have high similarity,which can be used as the training sample in the offline learning stage.(2)This paper focuses on the radar and infrared feature extraction and the selection of the optimal feature subset,and proposes a feature selection algorithm based on Owen value.After the feature extraction of radar and infrared respectively,combined with information theory and cooperative game theory,the cooperative game feature selection model is established.By calculating the Owen value of each feature,the radar optimal feature subset and infrared feature subset with high correlation with category,low redundancy between feature and feature,and strong dependence are selected from the original feature set of radar and infrared.From the two aspects of average identification accuracy and feature space evaluation,the simulation results show that compared with the other four comparison algorithms,the selected optimal feature subset still has the advantage of high separability in low dimension,which can effectively improve the target identification performance of information fusion system.(3)In this paper,the feature fusion of radar and infrared heterogeneous information is studied,and a feature fusion algorithm based on multi-core learning is proposed.In this algorithm,feature layer fusion and kernel layer fusion are combined to improve the accuracy.Firstly,the depth canonical correlation algorithm is used to project the feature information from the two types of detectors to the maximum correlation direction,then the weighted sum of the basic kernel function is carried out.The synthesis kernel is used to replace the single kernel function in the traditional classification algorithm,and the simple multi-core learning algorithm is used to train the classifier iteratively to determine the kernel function and its parameters suitable for the fusion of feature vectors.The experimental results show that the feature fusion algorithm based on multi-core learning can get higher identification accuracy by replacing the traditional single core learning with the synthetic core.