Research On Infrared Small Target Detection Using Spatial-Temporal Tensor Model

Infrared small target detection and recognition has been considered as the core and key technology of photoelectric detection system,as well as the research hotspot and difficulty in remote sensing science and photoelectric technology application field.Nowadays,with the increasing and expanding of national defense security and military requirements,infrared imaging detection and complex environment target detection technology has been more widely regarded.The existing detection algorithms pay insufficient attention to the temporal-domain information between the frames of the sequence image,the trajectory of the target,and the subtle dynamic changes of the background,so the detection effect is not good in the infrared sequence image with complex background.Aiming at the problems such as small targets,no obvious features,shape and size change with movement,and serious interference by background and clutter,this thesis focused on the efficient use of the spatial domain and temporal domain details to extract local prior information of target and background,combined with tensor recovery theory to improve detection efficiency,researched and proposed an infrared small target detection approach based on spatial-temporal tensor model.The problem of infrared dim small target detection is transformed into an optimization problem of the objective function.The main research contents are as follows:(1)The fundamental theories of tensor recovery are investigated.The constitution and slicing mode of tensor,expansion operation,norm definition,multiplication operation,and common tensor decomposition mode is introduced.Finally,the low-rank sparse decomposition model of the matrix is extended to the tensor recovery model,which provides a sufficient theoretical basis for the subsequent algorithm modeling.(2)The spatial-temporal model based on multi-frame is studied.Combining the tensor recovery model with multi-frame image detection,a spatial-temporal tensor model with the non-overlapping sequence is constructed to reduce invalid stacking operations.Finally,the three-dimensional tensor is extended to the four-dimensional tensor,and the4 D spatial-temporal tensor model is constructed for the first time,and its feasibility is proved by the singular value experiment.(3)A low contrast-constrained spatial-temporal tensor ring decomposition infrared target detection algorithm is proposed.Use local contrast to extract details.The tensor ring decomposition norm is introduced to carry out a convex approximation to the rank of the background tensor.While preserving the original structure of tensor data,the relationship between various dimensions during decomposition is reflected and the rank of the background is better represented.At the same time,the non-overlapping sequential spatial-temporal tensor is constructed.Finally,the ADMM framework is used to optimize the model.A large number of experiments and evaluation indicators prove that the model is excellent in algorithm robustness,background suppression ability,target enhancement ability,and computational efficiency.(4)A 4D spatial-temporal tensor infrared target detection algorithm based on the tensor train and tensor ring is proposed.Two efficient tensor network decomposition techniques are adopted: tensor train decomposition and tensor ring decomposition.4D spatial-temporal tensor is expanded into different matrix formats according to mode-(k)and mode-{n,l},and spatial-temporal features are extracted.The ADMM framework is used to optimize the model.The test results show that the model has better detection performance.Finally,the ablation experiment verifies that the dimension expansion from a 3D spatial-temporal tensor to a 4D spatial-temporal tensor can significantly improve the detection effect.