Research On Critical Technologies Of Long-range Imaging Recognition Of Near Ground Extended Targets

When the targets enter the detection range of optical system from millimeter wave radar,the 2D targets image information become the key of battlefield intelligence.However,optical transmission medias like turbulence,molecules,aerosols and other in the atmosphere that make the target images obtained by the near ground long-range imaging system fuzzy,geometric distortion,texture loss and other degradation.In order to improve the detection and recognition ability of those extended targets in the background of near ground long-range imaging,this dissertation firstly uses preprocessing methods to enhance the target images as well as remove the turbulence geometric distortion,then proposes an effective detection algorithm for moving-and-dim extended targets,and finally proposes an improved contour shape point set matching algorithm to improve the matching effection of the point set shape targets in the long-distance turbulence clutter scene.At the same time,the algorithm of fusing contour structure feature and heat kernal signature is proposed to realize the effective recognition of turbulent-distortaion contour target.The research contents are divided into five parts shown as follows:(1)An adaptive full scale Retinex enhancement(AFSR)method is proposed to enhance the energy enhancement.Different from traditional restoration methods with massive data and prior knowledge of natural imaging scenes to build mathematical models or deep network models,the proposed method can adaptively construct the full-scale surround function that guided by light transmission transmissivity.It improves the shortcomings of traditional Retinex method that can not represent the depth illumination information and manually adjusted scale parameters.At the same time,instead of logarithmic function operation,a simple linear approximation strategy greatly reduces the computational complexity of the algorithm,and a quasi real-time processing of0.055s/frame for 997×658×3 video image sequence is achieved.Lots of experiments show that the proposed algorithm can effectively improve the target image energy that obtained by the near ground diffraction imaging system,and also can enhance the discrimination of the extended target.(2)An adaptive mixed Gaussian subspace decomposition(AMoGSF)method is proposed to correct the geometric distortion of the turbulence-degraded images in the near ground long-range imaging scenes.In order to overcome the limitation of the adaptive optics(AO)system for the near ground anisoplanatic imaging,as well as the problem that the traditional(semi)blind restoration algorithm cannot obtain accurate prior knowledge and failing on the spatiotemporal variation distortion,a geometric distortion reduction algorithm without prior knowledge is proposed.The proposed method uses the mixture Gaussian distribution to model the components with target,noise and turbulence.Meanwhile,this method uses the online subspace decomposition strategy of multi-frame low rank structure to achieve the geometric distortion correction including the moving target.Lots of experimental results show that the proposed algorithm can improve the PSNR by about 30dB while reduce the normalized mean square error(NMSE)by about4.5%.This preprocessing methods improve the target identification ability and obtain a more complete contour shape extended target,which lays a foundation for the subsequent automatic target recognition.(3)In order to solve the problems that the camera jitter or rotation in airborne,shipborne and other near ground imaging scenes leading to the destruction of the low rank hypothesis of video sequence background,as well as the target recognition accuracy being reduced due to the intersection of turbulent motion and real object motion,a method named‘coarse to fine'for near ground long-distance moving target detection is proposed.In order to ensure the instability of low rank matrix caused by camera jitter or background movement,the T-AMoGSF model guided by transfer operator is proposed to ensure the low rank characteristic of dynamic background,and the robustness of the extracted'coarse'moving target in dynamic background is improved.At the same time,in view of the problem that it is difficult to detect the dim target when the turbulent motion and the real object motion are intertwined,the improved variable weighted pipeline filter(VWPF)method can make full use of the space-time structure of the multi frame sequence to effectively identify the dim target.The experimental results of seven methods in five different turbulence scenes are compared,and the experimental results show that the proposed method can achieve the best detection,the minimum missed detection rate and false alarm rate on the near ground visible and infrared imaging data sets.This method can directly detect and recognize the moving dim extended target of ROI of long-range imaging detection system.(4)An oriental shape context(OSC)energy cost function model with shape edge continuity constraint is proposed to improve the robustness of shape point matching recognition in turbulent clutter scenes.Firstly,an OSC description operator with rotation,scale and affine invariance is constructed.Then,a matching energy cost function model is constructed by shape edge continuity priori and the target cost function is optimized by using an ordered dynamic matching algorithm.The computational time complexity is reduced from the traditional O(n~4)to O(nm~2),where(n>m).Experimental results show that the proposed method is more robust than the other two typical methods,and the matching accuracy is improved by about 7%on average.(5)A fusion description operator is proposed to improve the shape context and geometric invariant features.In this method,a generalized shape feature fusion model is firstly derived,and then the sparse key points of the target contour shape in the turbulent clutter scene are effectively captured by the improved weighted discrete contour evolution algorithm(WDCE).On this basis,a fusion feature operator based on critical point inner distance shape context(CP-IDSC)and local scale invariant HKS(SI-HKS)is constructed for classification and matching recognition.Simulation and experiments on real turbulence data sets show that the proposed fusion feature descriptors are invariant to contour deformation,self occlusion,rotation and scale change,and can be effectively used for long-distance(infrared or visible imaging)targets recognition in turbulence degraded clutter scenes.Compared with other four typical methods,the proposed method has the highest recognition accuracy of 92.8%.In a word,this dissertation takes the problems as the guide,the algorithms model as the core,and the experiments as the foothold.Focusing on the key technologies of extended target intelligent recognition of near ground long-distance,some theoretical and engineering results has been achieved in this research work.