Infrared Dim And Small Target Detection Based On Low-rank And Sparse Recovery And Tensor Representation

Infrared thermal imaging system has the characteristics of high concealment,good anti-jamming,and strong adaptability to operating range and external environment,which plays an important role in modern military defense systems.It has been widely used in military detection,early warning,guidance and anti-missile,aerospace and other fields.For application scenarios that focus on concealment and generally have a long imaging distance,the imaging target is often called infrared dim and small target.This kind of target has a low SNR and occupies a small area,and also lacks some common features such as color,texture,shape,etc.,which causes difficulties with target detection.In addition to the characteristics of the target itself,the environment also increases the difficulty of detection.The radiation energy of the target captured by the system is already very weak,whereas the complex space environment may contain other high-intensity radiation sources,and the detection system itself also produce thermal radiation.How to detect infrared dim targets from complex and diverse backgrounds is a big challenge,which has attracted extensive attention from researchers around the world.The existing algorithms have poor robustness to clutter and noise,and the detection performance is unstable.Based on the framework of low-rank and sparse recovery and tensor representation,this paper studies the construction of tensor model of infrared image data and the characteristics of target and background.The main research contents are as follows:(1)The existing infrared dim target detection algorithms and related theories are studied.After analyzing the basic ideas and the merit and demerit of existing algorithms,two core research direction is determined,including low-rank and sparse theory and tensor representation;the basic concepts of tensor,tensor decomposition methods and other tensor related theorems are studied deeply.Moreover,the low-rank and sparse recovery theory are reviwed from two aspects of matrix representation and tensor representation.(2)A prior information based spatial-temporal tensor model for infrared small target detection is proposed.To take advantage of time information within sequential frames,the spatial-temporal tensor is constructed based on the property of the time and space correlation of background.Based on the prior information of the target and the structure tensor,an adaptive weight for the target is designed.The weight can effectively distinguish the sparse structures in the background from the target by the relative importance of the edge and corner.Based on the the prior information of the background,the non-local total variation is introduced to describe the piecewise smooth characteristic of the background and to reduce the interference to the target.Finally,the spatial-temporal tensor model based on edge and corner awareness is constructed,which improves the detection accuracy of the algorithm and the robustness to the clutter and noise.(3)A joint sparse regularized based twist tensor model for infrared small target detection is proposed.Based on the local continuity of the target in the direction of space and time,the twist tensor model is constructed by transforming the front view into the side view.The originally complex background components are more structured after the perspective conversion,and the difference between the background and the target increases.Based on this,the structured sparse inducing-norm is introduced to describe the locality and continuity of the target.In order to further suppress the interference of the background sparse structures and global noise,the structured sparse inducing-norm and the ₁l norm is combined and served as the joint sparse constraint of the target.The infrared small target detection model based on the joint sparse regularization and twist tensor can process single frame or sequential images with high detection accuracy and outstanding background suppression ability.