Research On Key Technology Of Imaging Seeker Object Recognition Based On Template Matching

Target recognition is one of the key technologies of the imaging seeker,and its performance directly affects the complex battlefield environment adaptability and strike accuracy of imaging guided weapons.In view of the requirements of imaging seekers for high-performance target recognition algorithms,this thesis studies the key technology of target recognition based on template matching.The purpose is to improve the ability of template matching algorithms to deal with differences in heterologous,rotation,scale,perspective and occlusion,and improve the accuracy and robustness of imaging seeker target recognition.The main work and achievements are as follows:(1)Aiming at reducing the search range of template matching,a fast localization method of initial target matching area based on the consistency of multi-source information is proposed.Firstly,the spatial consistency model of inertial/satellite information and seeker image information on the missile is established,and software compensation method is used to improve the time consistency between each information,so as to realize the prediction of the target position in the image.Secondly,according to the estimation of the maximum error of each information,the search range of the initial target matching is calculated and determined.Finally,experiments are carried out based on the real flight data of the imaging seeker.The results show that this method can effectively narrow the search range of target matching,and provide a good foundation for the implementation of the subsequent template matching algorithm.(2)Aiming at improving the robustness of template matching algorithms to deal with large differences in heterogeneity,rotation,scale,perspective,occlusion,etc.,a template matching algorithm based on Siamese network and target dense prediction is proposed.First,referring to the idea of the anchor-free fully convolutional Siamese network,the basic framework of the template matching algorithm based on deep learning is constructed.Then,on this basis,considering the need to improve the resolution of the output feature map and control the amount of computation,the feature extraction network is redesigned with reference to the idea of "encoding first and decoding later" in UNet,and the loss function is optimized.Thus,a template matching algorithm based on Siamese network and target dense prediction is formed.Finally,the experiment results on test data show that the template matching algorithm has good accuracy and robustness in complex environments.(3)To further improve the localization accuracy of template matching,a template matching algorithm based on Siamese network and center point estimation is proposed.Firstly,based on the research of template matching algorithm based on dense target prediction,the problem of predicting the whole target is transformed into the problem of estimating the target center with refer to the idea in Center Net.Then,a specific network and loss function for center point estimation are constructed,and a template matching algorithm based on Siamese network and center point estimation is formed.Finally,the qualitative and quantitative experiment results on test data show that the localization accuracy of template matching is effectively improved on the premise of maintaining robustness.(4)In order to solve the contradiction between the high computational complexity of the deep learning model and the limited computational resources of the seeker,a filter pruning algorithm that comprehensively considers the direct and indirect effects of filter pruning is proposed.Firstly,the limitation of existing methods in filter importance evaluation is analyzed.Then,a new method of importance evaluation that comprehensively considers the direct and indirect effects of filter pruning is proposed,which improves the accuracy of filter importance evaluation.Finally,based on the new importance evaluation method,a complete filter pruning algorithm is constructed.Experiments on public datasets show that this method can better maintain model accuracy when the model is greatly compressed.