Infrared Small Target Detection Via Tensor Recovery

With the continuous upgrade of global arms competition,infrared small target detection technology is indispensable in precision guidance,infrared warning and other systems.Because infrared images often have low contrast,coupled with complex natural backgrounds and long imaging distances,infrared small target detection is full of challenges.It is of great significance to propose efficient,robust and real-time algorithms.At present,the existing algorithms exist several drawbacks.First,it's difficult to balance time and performance;second,the robustness against variety scenes is not high;third,the local and global features are not combined;fourth,edges can not be well suppressed;and fifth,the anti-noise ability is not strong.As an extension of the correlation matrix framework,tensor recovery theory has become one of the research hotspots in the field of image processing in recent years.In this thesis,based on this new method,by constructing the three-dimensional tensor representation of single infrared image,the background and target could be separated by solving the preset tensor recovery model.The specific research contents are as follows:(1)The low-rank matrix recovery theory and the infrared small target detection methods based on this theory are studied.The basic framework of low-rank matrix recovery theory and its commonly used solutions are introduced.The detection models of infrared small target based on this theory are fully analyzed,advantages and disadvantages of which are summarized at the same time.Finally,the advantages of tensor modeling are pointed out.(2)The tensor recovery theory is studied in this thesis.As the core theoretical tool,the following aspects of tensor recovery theory are studied in depth,including the basic concepts of tensor,operations between tensors,common tensor decomposition methods,tensor recovery models and their corresponding solutions.(3)Infrared small target detection based on partial sum of the tensor nuclear norm is proposed.To improve the scene robustness and reduce the running time,inspired by the infrared patch tensor model,this topic starts with studying the extraction method of local features and the selection method of tensor rank metrics.By combining the local and non-local features,the scene robustness is greatly enhanced,and the running time is dramatically reduced.(4)Infrared small target detection based on stable principle component pursuit with three-dimentional anisotropic total variation is proposed.The approaches to enhance the noise robustness and edge suppression ability of the model are in-depth researched.By introducing a three-dimentional anisotropic total variation term,a joint constraint of total variation and low rank is applied to the infrared background,which can more accurately describe the changes in gray levels in the background,thereby improving the performance of small target detection.In addition,replacing the traditional model with the stable principal component pursuit model increases the robustness of the algorithm to noise.