Research On Dim And Small Target Detection Algorithm In Spatio-temporal Filtering And Background Modeling

In order to improve the detection performance of the detection system for dim and small targets in non-stationary scenes,this dissertation focuses on the key technologies of dim and small target detection.Firstly,background suppression is performed on two types of non-stationary scenes encountered in photoelectric imaging detection to improve the contrast between target and background.Then,the mutual enhancement of target energy is performed on the difference images obtained by background suppression.Finally,the improved target detection algorithm is used to detect and extract the enhanced targets.The main research works are as follows:(1)A low-rank background modeling algorithm(CLRA)based on statistical clustering partition is proposed for multi-component non-stationary complex scenes.In order to effectively suppress the clutter textures caused by the undulating and non-stationary backgrounds to reduce the false alarm sources in the difference images,improve the contrast of the targets,CLRA algorithm uses the statistical characteristics of images to perform adaptive clustering and obtain more stable image regions,which can effectively regional stabilize the non-stationary scene images,achieve more accurate low-rank background estimation,and achieve the difference images with less residual,lower value and more evenly distributed noise.Therefore,the influence of undulating non-stationary components on the target is effectively suppressed.The experimental result data of three representative complex scenes are as follows: The mean values of structural similarity are 0.9735,0.8357 and 0.9776,the mean values of noise floor are 8.7269,13.9324 and 11.6636,and the mean values of contrast are10.4446,4.7972 and 6.4491,respectively.Experimental data indicates that CLRA algorithm achieves significant background suppression effect: residual textures are effectively removed in the obtained difference images,the distribution of noise is uniform,and the targets are highlighted.(2)A single-pixel background modeling algorithm(SPA)based on the spatial basis function of local region background is proposed for the non-stationary scene with strong vignetting(strong sky scene).The strong vignetting in the strong sky scene has the characteristics of random distribution,high intensity,wide range,large dynamic fluctuation and often drowning the target.Therefore,in order to achieve extremely weak target detection performance,SPA algorithm uses pixel neighborhood backgrounds to represent its spatial basis function,and transforms the modeling of strong vignetting background into an optimal estimation problem to obtain the optimization result of further balance between background components and sparse components.Thus,the extremely weak targets can be effectively separated from the strong undulating backgrounds.Experimental data of three typical strong sky scenes indicates that SPA algorithm achieves good background suppression and target enhanced performance,which is mainly reflected in the following three aspects: Firstly,it has good robust performance for targets with different SNR,velocity and trajectory characteristics.Experimental results show that the algorithm can effectively separate the background and target components for small targets with SNR less than or equal to 1.5d B,moving speed up to 2 pixels/frame and complex trajectory.Secondly,residual noise has the characteristic of white noise.The residual noise of the difference images obtained by the algorithm is evenly distributed and has a spectrum similar to that of white noise.Thirdly,targets have been significantly enhanced.The mean values of SNR gains obtained for the three scenes are 6.7709 d B,6.7812 d B,6.6343 d B,respectively.The mean values of background suppress factor are 2.8061,1.7039 and 13.8459,respectively.(3)Adaptive spatial-temporal energy mutual enhancement pipeline filtering algorithm(ASEMEPF)is proposed to detect dim and small targets in the case of strong noise and targets are temporary disappeared(to be obscured or submerged).The ASEMEPF algorithm is composed of adaptive spatial-temporal energy mutual enhancement algorithm(ASEME)and adaptive pipeline filtering target detection algorithm(APFA).In order to further suppress the non-target components of difference images obtained by CLRA and SPA algorithms in the early stage,reduce false alarm sources and enhance target energy,so as to facilitate stable and effective target detection in the later stage,ASEME algorithm is proposed.The algorithm constructs an adaptive gradient reciprocal filter based on the difference of distribution characteristics and statistical characteristics between non-target and target components to further suppress the non-target component and realize mutual enhancement of target energy.Experimental data of four representative scenes show that the contrast mean values of the targets are increased by 28.41,22.44,47.54 and 66.48 respectively compared with the original difference images after energy enhancement by ASEME algorithm,indicating that the targets have been significantly enhanced.Following,in the process of target detection,in order to improve the robustness of pipeline filtering algorithm against strong noise and targets are temporary disappeared,the APFA algorithm based on Kalman filter is proposed.The algorithm realizes the target information compensation and the pipeline center and diameter adjustment adaptively by combining the target prediction information with the existing target motion information and gray information.Experimental data indicate that the whole ASEMEPF algorithm achieves good target detection performance for the four representative scenes: the target detection rates are 100% or approaching 100%,and the error detection rates,missed detection rates and false alarm rates are zero or approaching zero.The temporarily disappeared targets can be estimated effectively and the target detection trajectories are complete and smooth.