A Robust Tensor Decomposition Model For Infrared Small Target Detection Based On Bayesian Framework

The development of infrared imaging technology accelerates the research and application in infrared image processing.Among them,the research of infrared small target detection is a hot issue in the field of military,transportation,monitoring and so on.In this paper,a small target detection method based on Bayesian model is proposed to solve the problem using airborne infrared small target image sequence.In order to maintain the integrity of the data during the decomposition process,and reduce the damage of high order data dimensionality reduction,a new method using higher order tensors decomposition is proposed to replace the traditional method of two-dimensional data decomposition in small target detection.Based on the advantage of spatial constraint of 3D data,the single target to multi-target detection is realized.The content of the study is divided into the following three aspects.Firstly,We introduce the basic content and application of tensor and tensor decomposition.Then two kinds of tensor decomposition methods are studied,and the similarities and differences of the two kinds of algorithms are compared.The two methods are introduced in detail from basic concepts to practical examples.So we pave the way for the following tensor applications.Secondly,the model of robust incomplete tensor decomposition under Bayesian framework is established,which is a probabilistic model.There are three problems in the probabilistic model of robustness and incomplete tensor decomposition.First,the probabilistic model selects the complex prior distribution,resulting in slow convergence.Second,we need to specify the rank of the tensor before the tensor decomposition,and the rank of the tensor is a difficult NP-hard problem.Third,the use of adjustable parameters to establish the prediction model,resulting in it ' s unsteady.Our model is improved to solve the above problems.In this paper,the process of establishing the hierarchical probability model is introduced in detail,and the principle of automatic inference of CP rank is illustrated.After the model is built,the algorithm is applied to the field of infrared small target detection for the first time.Choosing different small targets such as ground,sea and sky as input to carry out experiments.We change the infrared image sequence into a natural length,width,time dimension on the order of three tensors as input,then the CP decomposition method is used to decompose it.The result of decomposition is composed of three parts: lowrank,sparse and noise.Among them,the sparse part as the foreground module target detection results.The experimental results show that the model can be applied to single target and multi-target detection of infrared small target.Finally,Quantitative analysis of the performance of this model,compared with the traditional infrared small target detection algorithm.The feasibility and superiority of the algorithm in infrared small target detection is demonstrated by the combination of subjective observation and objective analysis.Several evaluation indexes are defined to test the performance of the algorithm in the infrared small target detection based on tensor decomposition.