Search

CN-116664865-B - GoDec-based infrared small target detection method and GoDec-based infrared small target detection device

CN116664865BCN 116664865 BCN116664865 BCN 116664865BCN-116664865-B

Abstract

The invention discloses a GoDec-based infrared small target detection method and device. Comprises (1) constructing an infrared space-time tensor from a sequence of thermal infrared images And Expanded in a time dimension to obtain a reconstruction matrix (2) Yes, that's right Performing singular value decomposition, determining low rank constraint of background component, (3) based on assumption of infrared image composed of low rank background component and non-background component, aiming at minimizing Welsch M estimation of non-background component, establishing GoDec-based decomposition model, (4) designing optimization solution algorithm to obtain matrix of non-background component (5) Designing an information filter to remove non-data Noise in the matrix to obtain pure sparse target matrix And Reconstructing the image matrix T with the same size as the original infrared image to obtain a target detection result sequence T, and realizing thermal infrared small target detection. The invention can effectively inhibit the background, strengthen the target and improve the detection performance of the infrared small target.

Inventors

  • LI XIAORUN
  • LUO YUAN
  • CHEN SHUHAN
  • WANG JING

Assignees

  • 浙江大学

Dates

Publication Date
20260505
Application Date
20230518

Claims (7)

  1. 1. The GoDec-based infrared small target detection method is characterized by comprising the following steps of: step 1) for an original thermal infrared image sequence, stacking successive image frames to construct an infrared spatio-temporal tensor And tensor the space-time Expanding in time dimension to obtain a new reconstruction matrix Step 2) reconstruction matrix Performing singular value decomposition, and obtaining the rank r of the background component in a self-adaptive manner through the obtained singular value matrix; step 3) with the minimum Welsch M estimation non-background component as a target, utilizing the rank r in the step 2) to establish a matrix decomposition model based on GoDec; The step 3) is specifically as follows: the GoDec algorithm is modified to reconstruct the matrix As an object, based on a reconstruction matrix The assumption consisting of low rank background components and non-background components, with minimized Welsch M estimated non-background components as the objective function, the low rank sparse tensor decomposition model is improved to the matrix decomposition model as follows: Wherein, the Representing a matrix of low rank background components, Representing a matrix of non-background components, r being the rank of the background component determined in step 2), Representing a matrix of non-background components G σ (·) represents a gaussian kernel function with σ as standard deviation, and rank () represents a rank calculation operator; Step 4) designing a model optimization solving algorithm based on a quadratic optimization theory, and solving a matrix decomposition model based on GoDec by using an alternate solving method to obtain a non-background component matrix Step 5) designing an information filter by removing the matrix of non-background components Screening out candidate target pixel points to obtain a pure target matrix And matrix the object Reconstructing the image matrix T with the same size as the original infrared image to obtain an infrared small target detection result sequence, and realizing thermal infrared small target detection.
  2. 2. The method for detecting small infrared targets based on GoDec according to claim 1, wherein the step 1) specifically includes: for each frame of image in the original thermal infrared image sequence I 1 and I 2 respectively represent the length and width of an image frame, and each image frame is modeled as consisting of a low-rank background matrix B, a sparse target matrix T and a noise matrix N, and the expression is that Wherein rank (·) represents the rank calculation operator, and 0 represents l 0 norm, F represents the Frobenius norm, λ 1 and λ 2 represent positive trade-off parameters; From the original thermal infrared image sequence, the number of each time of acquisition of continuous image frames is I 3 , and the image frames are sequentially stacked according to the frame sequence of the image frames to construct a space-time tensor And The s-th front slice of (2) is the s-th complete image frame, thereby modeling the space-time tensor as a low-rank sparse tensor decomposition model expressed as Wherein, the The background tensor is represented as such, The target tensor is represented by a set of parameters, Representing a noise tensor; Space-time tensor Each tube fiber of (3) Carrying time information, wherein I is more than or equal to 1 and less than or equal to I 1 ,1≤j≤I 2 , and tensor of the time and the space is calculated Expanding in time dimension to obtain a new reconstruction matrix The reconstruction matrix The first dimension of (2) carries temporal information, the second dimension carries image space information, and each row vector of the matrix is reconstructed Represents an infrared image sample, wherein x is equal to or more than 1 and is equal to or less than I 3 .
  3. 3. The method for detecting small infrared targets based on GoDec according to claim 1, wherein the step 2) specifically includes: Reconstruction matrix Singular value decomposition is performed, and the expression is as follows: Wherein, "x" is a matrix multiplication operator, U is a left singular matrix with the size of I 3 ×I 3 , V is a right singular matrix with the size of I 1 I 2 ×I 1 I 2 , Σ is a singular value matrix with the size of I 3 ×I 1 I 2 , and V T represents the transposition of the matrix V, I 1 and I 2 respectively represent the length and the width of an image frame, and I 3 is the number of continuous image frames obtained each time; the method for determining the rank r of the background component is as follows: Wherein H (·) is a unit step function, when x is equal to or greater than 0, H (x) =1, otherwise H (x) =0, Σ k,k represents the k-th row, k-th column element in the singular value matrix Σ, Σ max represents the maximum value element in Σ, and r Σ represents an adjustable positive constant.
  4. 4. The method for detecting small infrared targets based on GoDec according to claim 1, wherein the step 4) specifically includes: based on the theory of quadratic optimization, by introducing The auxiliary variable E of (2) is equivalent to the matrix decomposition model in the step 3) Wherein phi (-) is the dual function of Gaussian kernel function g σ (-), and further solving formula (6) by an alternate solution method, wherein the obtained alternate solution model is Wherein t is the number of times of iteration at the t th time; finally, obtaining a low-rank background component matrix by alternately solving the model And a non-background component matrix
  5. 5. The method for detecting small infrared targets based on GoDec according to claim 4, wherein the specific solution algorithm of the alternative solution model is: 1) Input reconstruction matrix Error epsilon, rank r of the background component determined in step 2), iteration number t=1; 2) Random initialization standard Gaussian matrix 3) Calculation of When (when) The following steps 4) -9) are performed), whereas step 10) is performed; 4) 5) 6) Performing QR-decomposition on matrix Y 2 , Y 2 =qr, and k=q; 7) wherein, the root of Hadamard product; 8) The number of iterations t=t+1; 9) Returning to the step 3); 10 Output background component matrix
  6. 6. The method for detecting small infrared targets based on GoDec according to claim 1, wherein the step 5) specifically includes: The non-background component matrix resulting from step 4) In each row Representing non-background part in each infrared image, including noise component and target component, and because the pixel gray value of target component is greater than noise component, constructing information filter to screen out target candidate pixel point, removing noise component to obtain pure sparse target matrix The information filter is designed as follows: Wherein, the Representing a row vector The element whose mid index value belongs to the set Ω x projects onto itself, the operator that projects all other elements to zero, Representing the acquisition of the x-th row vector The index set omega x corresponding to the element with the large middle and front kappa x is an operator, kappa x is eta I 1 I 2 , eta is an adjustable positive constant, eta is epsilon (0, 1), I 1 and I 2 are the length and the width of an image frame respectively, and I 3 is the number of continuous image frames obtained each time; further, the pure target matrix Is defined by each row vector of (a) Reconstructing an image matrix T with the same size as the original thermal infrared image, wherein x is more than or equal to 1 and less than or equal to I 3 , and obtaining I 3 images, thereby obtaining an infrared small target detection result sequence and realizing infrared small target detection.
  7. 7. A GoDec-based infrared small-target detection device embodying the method of claim 1, comprising: the infrared image reconstruction module is used for constructing an original thermal infrared image sequence into a space-time tensor and expanding the space-time tensor into an image matrix in a time dimension; A low-rank constraint module for determining a background component, the low-rank constraint module being used for determining a low-rank constraint of the background component in a subsequent GoDec-based decomposition model; a matrix decomposition model building module based on GoDec, which builds a matrix decomposition model based on the assumption that the infrared image consists of low-rank background components and non-background components and taking minimized Welsch M estimated non-background components as an objective function; The optimization model solving module is used for solving a matrix decomposition model based on a quadratic optimization theory through an alternate solution method to obtain a non-background component matrix; The information filter is used for removing noise components in the non-background component matrix, screening out target candidate pixel points to obtain a pure sparse target matrix, and reconstructing each row of the target matrix into an image matrix with the same size as the original infrared image to obtain an infrared small target detection result sequence; and the target detection result output module is used for outputting an infrared small target detection result graph.

Description

GoDec-based infrared small target detection method and GoDec-based infrared small target detection device Technical Field The invention relates to the field of image processing, in particular to a GoDec-based infrared small target detection method and device. Background The detection of the thermal infrared small target plays an important role in the fields of military reconnaissance, border patrol, security monitoring, environmental monitoring and the like. In general, small infrared targets exhibit small size, lack of texture features, weak signal strength, and infrared images of poor quality and low signal-to-noise ratio make the task of small target detection more difficult. Therefore, the research of the infrared small target detection algorithm is of great significance. At present, the thermal infrared small target detection method mainly comprises a traditional method, a method based on visual saliency, a method based on low-rank sparse decomposition and a deep learning method. Classical traditional methods include background modeling-based methods, wavelet transform-based methods, and the like, which can extract small target features in thermal infrared images, but detection results are easily interfered by noise and background and perform poorly in complex and diverse scenes. Inspired by the human visual system, given that infrared small targets typically exhibit local features that are different from background and noise, infrared small targets can be detected by designing specific local descriptor operators, common algorithms are local contrast metric LCM operators and their various variants, such as multi-scale block contrast metric MPCM, weighted local contrast metric WLCM, weighted multi-derivative based contrast metric MDWCM, and so on. However, these detectors are highly dependent on the contrast of the target with the background, and are prone to false positives when a small target is more similar to its surrounding neighborhood. The low-rank sparse decomposition theory injects new vitality for infrared small target detection. Since most of the area in the infrared image is occupied by the background component of the local autocorrelation, the background exhibits a low rank characteristic, while the object occupies only a few pixels, and thus the object exhibits a sparse characteristic. Therefore, the modeling can be performed on the low-rank background and the sparse target, the target components are separated from the original infrared image, and the detection of the infrared weak and small target is realized. Common methods are GoDec, IPI, NRAM, logTFNN and RIPT, etc., which show better performance in preserving target information and suppressing background and noise. However, many methods do not take full advantage of, or even destroy, the temporal and spatial information of the thermal infrared image sequence, which may lead to background shrinkage or target loss problems. Along with the development of deep learning technology, researchers put forward many thermal infrared small target detection methods based on deep learning, for example ALCNet, CBPNet, STDMANet and APAFNet, but these methods based on deep learning require a large amount of labeling data and high-performance hardware equipment, and in addition, phenomena such as feature loss, target misjudgment, weaker generalization capability and the like easily occur in the network training process, so that the robustness of an algorithm is reduced under the condition of severe background environment change, and therefore, the method based on deep learning still has a certain limitation in the practical application process. Disclosure of Invention Aiming at the defects in the prior art, the invention aims to provide a thermal infrared small target detection method and device based on GoDec, which fully utilize the space-time information of a thermal infrared image sequence, the low-rank characteristic of background components and the sparse characteristic of targets, organically combine a maximized related entropy criterion with a low-rank sparse decomposition theory, combine an adaptive low-rank background constraint and an information filter, and establish a component decomposition model based on a low-rank background part and a non-background part, wherein the model can effectively represent the characteristics of different components of infrared image data. In addition, based on a quadratic optimization theory, an efficient optimization solving algorithm is provided, the thermal infrared small target can be detected rapidly, the infrared small target detection is realized, and the target detection capability and the background inhibition capability of the algorithm are verified effectively. In order to achieve the above purpose, the present invention provides the following technical solutions: The invention discloses a GoDec-based infrared small target detection method, which is characterized by comprising the following steps o