Research On Infrared Small Target Detection Method Based On RPCA

With the rapid development of infrared imaging technology,infrared imaging has been widely used in military,aerospace projects due to its excellent imaging capabilities.Although infrared images have strong anti-interference ability and high environmental adaptability,the difficulty of detecting small targets in infrared images increases sharply due to the long imaging distance of infrared images.Small objects in infrared images lack texture information and spatial structure information,and only occupy a small part of the image,which makes the detection task very difficult.To make matters worse,the background of the infrared image is very complex,and the clutter and strong edges in the background are often mistaken for the target,thus increasing the false alarm rate.The existing infrared small target detection methods often cannot balance the relationship between suppressing the background and enhancing the target.Therefore,this thesis proposes a new infrared small target detection method based on the RPCA method.(1)In order to make full use of the background information of the image and improve the accuracy of small object detection,we turn the object detection problem into a low-rank sparse tensor optimization problem.In order to improve the low rank of the matrix,some non-convex functions on the singular values of the matrix are proposed for matrix completion.Therefore,we apply non-convex functions to squaring matrices for background matrix completion and extend to the tensor domain.This method provides stable performance and high computational efficiency from schema construction.Based on a non-convex relaxed low-rank tensor completion method,we propose in this thesis to approximate the rank of the background tensor using a non-convex function on the singular values of the squaring processing matrix of the tensor.(2)We found that under the multi-angle derivative,the target has stronger edge features than the background clutter,but the commonly used method only takes the difference between the eigenvalues of the structure tensor,which can easily obtain the edge of the background but at the same time the target The edges are also highlighted,causing excessive shrinkage of the target.We propose a prior analysis based on multiple angles,coupled with reweighting,to ensure that stubborn background edges are removed without shrinking small objects in the past.(3)In this thesis,we propose a non-convex relaxation-based low-order tensor completion infrared small target detection model based on the IPT model,and use an efficient algorithm based on multiplier alternating directions to solve the model.In addition,our model has also achieved good results in comparison experiments with some advanced algorithms.