Research On Infrared Dim And Small Target Detection Theory And Methodology Based On Sparse Dynamic Inversion

Infrared imaging and detection has been widely used in infrared guidance and antimissile system,infrared early warning system,long-distance detection and space debris detection,etc.Target detection and intelligent recognition are necessary in such applications.Due to the long imaging distance,complex imaging situation and the quality of infrared sensors,the detection of infrared small targets has become a challenging problem,which draws wide attention.Traditional infrared small target detection methods have their drawbacks,such as the unstable performance against background clutter and noise,which is not suitable for extensive usage in practical.Sparse dynamic inversion,or sparse dynamic optimization,is a common approach in signal processing used for extracting certain components from input signals,which could further reveal important information.By analyzing infrared imaging background and infrared small targets,infrared small target detection performance could be improved,combining the sparse dynamic inversion theory.The stability and the effectiveness of infrared search and tracking systems would also benefit from this research.The content of this thesis focuses on the analysis of infrared images,model construction,sparse dynamic inversion approach,etc.The scenes include sea background,sky background and ground background,with various distribution and complexity.The main contents are listed as follows:(1)The infrared imaging background and target features are fully analyzed.The characteristics of different backgrounds and the difficulties of detection tasks are cleared.Traditional infrared small target detection methods didn't consider much about the diversity of backgrounds and infrared targets,which leads to the instability under clutter scenes.In this thesis,by starting from the infrared imaging theory and system,the solid foundation of target/background model construction is build,which also serves as the theoretical base of the research work in this thesis.(2)An infrared small target detection approach based on joint regularization is proposed.The principal component pursuit(PCP)model is combined with the total variation(TV)regularization method,named as TV-PCP method.In infrared scenes,due to the blurring features and the lack of detailed information,the infrared background could be considered as low-rank,or partly low-rank.Under this assumption,the edges within background component could be the interruption for target detection.By introducing the total variation term into the PCP model,the changing components in background can be well described,which prevents it from interrupting the target detection.By building the TVPCP joint regularization model,the performance of target detection as well as background estimation has been improved.(3)An infrared small target detection method based on stable multi-subspace learning is proposed.Multiple infrared radiation sources could exist in complex infrared background,which would interfere the detection of infrared small targets.The scenes with multiple radiatio sources were not fully considered in traditional infrared small target detection methods.In this thesis,we propose to use the multi-subspace learning based methods to deal with highly complex background.Different from the single subspace learning,multiple subspace learning is more suitable for dealing with the afore-mentioned complex scenes.The details in background can be well modelled,which would help with the detection of infrared small targets.The proposed method has been proved to be effective.(4)A multi-task joint sparse representation based infrared small target detection approach is proposed.Because of the lack of information in infrared images,sparse representation-based detection approaches always face difficulties in the construction of samples and the choice of features.In practical,it would lead to false detection or missed detection.We propose to use a novel multi-task joint sparse representation and classification method to build the infrared small target detection system.Four features are used,including grey feature,edge feature,histogram of gradient feature and grey histogram,to classify all samples into three classes: target class,smooth background class and changing background class.An optimization solver based on smoothed L0 norm is also introduced in the proposed approach.Experimental results show good target detection performance of the proposed method.(5)A non-negativity constrained variational mode decomposition method is proposed.Different components in infrared images,including background,targets and noise,they have different grey level distribution in spatial domain.As a result,the frequency distribution is also different.Mode decomposition is a common approach for signal decomposition,which is widely used for decomposing input signal into sub-signals with different features.An adaptive narrow-band signal decomposition method,variational mode decomposition(VMD),is used for infrared small target detection.Considering the non-negativity feature of infrares bright targets,the corresponding constraint is introduced to the VMD method.The pre-processed infrared image is adaptively decomposed into narrow-band sub-signals.The infrared small target is distributed in one of the sub-signals.By this means,the proposed method is computational efficient,with small amount of clutters in the detection results,as well as with low false alarm rate.It outreaches the existing methods under certain types of infrared background.(6)Two visual-saliency based infrared small target detection methods are proposed,using two different ways to detect the infrared small targets.One of the approaches uses the local contrast of infrared small targets.The potential target area is located by using local entropy and the local similarity of infrared images.After that,the local contrast is computed in the potential area,which is a highly efficient approach.Another saliency detection approach is based on the grey saliency and the motion saliency of infrared small targets.A boolean map saliency method combined with motion feature is proposed.By computing the boolean map saliency as well as the motion saliency,the infrared small target can be easily located,especially in dynamic scenes.Experiments results on real dataset and the evaluation results show that,the proposed approaches in this thesis solve the problem of infrared small detection with low false alarm rate and high detection rate.It also solves the problem of missed detection in some scenes.Combined with the infrared background analysis we did in this thesis,the proposed approaches are well designed for certain backgrounds,which can be further applied in infrared search and track systems.Additionally,we build a theoretical system of infrared small target detection based on sparse dynamic inversion,which is a good starting point for the follow-up applications.