Infrared small target detection method based on structured background tensor estimation method

Abstract

The application discloses an infrared small target detection method based on structured background tensor estimation, and belongs to the field of infrared image small target detection. The main steps are as follows: an original infrared small target image is divided into a target tensor and a background tensor, the background tensor is modeled, and a linear process is used to simulate a real background degradation process; a low-rank learning model of structured background tensor estimation is constructed, a subspace tensor and a coefficient tensor obtained by structured decomposition of the background tensor are subjected to regularization constraint, and a core coefficient tensor is decomposed and inverted in a cascaded manner; an alternating direction multiplier method is used to optimize the obtained formula; and the whole model is trained according to the solution of the subspace and the coefficient tensor, so that the infrared small target detection is realized. By using the application, the background information of the image can be fully utilized, and the precision of small target detection is improved.

Description

Technical Field

[0001] This invention belongs to the field of infrared small target detection, specifically, it is an infrared small target detection method based on the structured background tensor estimation method. Background Technology

[0002] Infrared small target detection has become a highly regarded research area in recent years, with applications spanning military, aerospace, security monitoring, and environmental monitoring. With the continuous development and widespread adoption of infrared imaging technology, the quality and resolution of infrared images have significantly improved, providing more possibilities for infrared small target detection. Simultaneously, deep learning technology has made significant progress in computer vision, especially the application of convolutional neural networks in target detection, which has become a research hotspot. However, due to the long imaging distance of infrared images, infrared small target detection is often affected by various interferences, such as thick cloud edges, sea surface clutter, and artificial ground heat sources, making target detection extremely difficult. Furthermore, infrared images have complex backgrounds and high imaging noise, typically resulting in a very low signal-to-noise ratio. Small targets lack texture and spatial structure information, occupying only a small portion of the image. Therefore, efficiently separating background clutter from the target is a major challenge in infrared small target detection.

[0003] Traditional infrared small target detection methods mainly include the following: Threshold-based methods: By setting an appropriate threshold, pixels with higher brightness in the infrared image are identified as targets. This method is simple and direct, but it is easily interfered with by changes in lighting and complex backgrounds; Motion detection methods: These methods utilize the motion information of the target in consecutive frames for detection. Common techniques include optical flow and inter-frame difference methods. This method performs well in detecting moving targets, but poorly in detecting stationary or slow-moving targets; Morphological methods: These methods use morphological operations to extract targets from infrared images, such as using opening and closing operations to eliminate noise and connect target regions; Filter-based methods: These methods use correlation filters to process the infrared image and utilize the correlation between the target and the template for target detection; Feature matching methods: These methods extract features from the infrared image and perform target detection through feature matching. Common features include grayscale features and texture features. The paper "Infrared Patch-Image Model for Small Target Detection in a Single Image" proposes an IPI infrared model that converts the original image into a matrix. By using a sliding window to overlay partially overlapping image patches, the correlation with the background is increased. The sparse and low-rank matrices are recovered from the original matrix. This model transforms the target detection problem into an optimization problem of recovering the low-rank and sparse matrix of the background, ultimately achieving separation between the target and the background. Han et al. proposed a method based on relative local contrast measurement (RLCM), which uses relative local contrast measurement to enhance the real target. Zhang et al. proposed a detection model based on self-regularized weighted sparsity (SRWS), which uses overlapping edge information from background structure information to constrain the sparse term. Zhang et al. proposed a model based on... The norm-constrained nonconvex optimization detection method (NOLC) is used to... Norms strengthen the constraints on sparse terms, transforming the NP-hard problem into a non-convex optimization problem. While the aforementioned model considers background-related prior issues and constraints, it still suffers from problems such as excessive target shrinkage, clutter residuals, and long processing times. In conclusion, the capabilities of existing infrared small target detection algorithms need improvement. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides an infrared small target detection method based on a structured background tensor estimation approach. This method primarily involves modeling the background tensor, utilizing a subspace low-rank learning model and minimizing the cascaded rank of the coefficient tensor to fully leverage the background information of the image, separating background clutter from the target. This satisfies the need for small target detection in infrared images with background interference, thereby reducing the false alarm rate.

[0005] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows: An infrared small target detection method based on structured background tensor estimation includes the following steps: An infrared small target detection method based on structured background tensor estimation, characterized in that the method includes the following steps: S1. Divide the original image tensor into target tensors. Background tensor and noise tensor ; S2. Model the background tensor using a linear process. A degradation model is obtained by simulating the real background degradation process; S3. Construct a low-rank learning model for structured background tensor estimation, projecting the background tensor onto a set of base values. In the resulting subspace, the degenerate model is transformed into a regularized tensor estimation model; S4. Coefficient tensor obtained from background tensor decomposition Regularization constraints are applied to decompose and invert the core tensor in a cascaded manner to remove the spatial redundancy information of the degenerate tensor. S5, Subspace and coefficient tensor The constraint terms are added to the structured background tensor estimation model, and the formula is optimized using the alternating direction multiplier method. S6. Optimize and solve the parameters of the structured background tensor estimation model constructed in step S5, iteratively train the structured background tensor estimation model, and complete the infrared small target detection.

[0006] Furthermore, the original image tensor can be described as:

[0007] Furthermore, the background tensor for infrared small target estimation contains a large amount of noise caused by clouds, sea clutter, etc., resulting in highly redundant spatial information. This redundant spatial information leads to severe degradation of the original infrared background image. To mimic the real background degradation process, we not only consider existing work (directly modeling the background tensor) but also a linear process. By simulating the real background degradation process using a linear process, the resulting degradation model is as follows:

[0008] Furthermore, in step S3, the background tensor Decompose into base and coefficient tensor , According to the maximum a posteriori estimation theory, the degenerate tensor model is transformed into a regularized tensor estimation model: , in, It is a coefficient tensor Constraints, It is a regularization parameter. Represents the tensor modulus product, It is the Frobenius norm used as a regularization term in image restoration problems.

[0009] Furthermore, due to factors such as imaging distance and environment, small infrared targets lack texture and spatial structure information, occupying only a small portion of the image. Therefore, the target can be considered as a sparse matrix. In contrast, many local patches in the background image are approximately linearly correlated, so the background tensor... This can be viewed as a low-rank matrix. Therefore, the infrared small target detection problem is transformed into: .

[0010] Furthermore, the coefficient tensor obtained from the background tensor decomposition The coefficient tensor contains redundant spatial information, therefore the coefficient tensor... Regularization constraints are applied, and tensor structured cascaded rank decomposition is used to effectively utilize... The inherent correlation and low rank of the core tensor By decomposing and inverting in a cascaded manner, the spatial redundancy information of the degenerate tensor is removed, thus the coefficient tensor... constraint terms It can be represented as:

[0011] in It is a linear transformation operator. Represents the coefficient tensor slices, Represents the core tensor. Indicates to The proposed method minimizes the rank of tensor cascades.

[0012] Furthermore, step S5 will process the subspace and coefficient tensor Adding the constraint terms to the model of the background tensor yields:

[0013] Furthermore, the Alternating Direction Multiplier Method (ADMM) algorithm is used to solve the model:

[0014] Compared with the prior art, the technical solution of the present invention has the following technical effects: This invention discloses an infrared small target detection method based on structured background tensor estimation. It models the separated background tensor and uses a linear process to simulate the real background degradation process. By fully utilizing the background information of the image, a low-rank learning model for structured background tensor estimation is constructed. Constraints are applied to the subspace tensor and coefficient tensor obtained from background tensor decomposition, overcoming the drawback of simultaneously minimizing the nuclear norm and constraining all singular values. By utilizing the structured background tensor estimation model and the cascaded rank minimization of the coefficient tensor, background clutter is efficiently separated from the target, improving the accuracy of infrared small target detection. Attached Figure Description

[0015] Figure 1 This is the algorithm flowchart of the present invention.

[0016] Figure 2 This is a diagram illustrating the overall framework of the model of this invention.

[0017] Figure 3 This is the original infrared image of the present invention.

[0018] Figure 4 This is a 3D image of the target result image and the result image after detection by this invention.

[0019] Figure 5 It is a 3D diagram of the target result map and the result map of the IPI method.

[0020] Figure 6 It is a 3D diagram of the target result map and the result map of the RLCM method.

[0021] Figure 7 These are 3D diagrams of the target result map and the result map of the SRWS method.

[0022] Figure 8 These are 3D plots of the target result map and the result map of the NOLC method. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to specific embodiments and corresponding drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0024] Example: Figure 1 As shown, an infrared small target detection method based on structured background tensor estimation is described, with the following specific steps: S1. Divide the original infrared image tensor into a target tensor. Background tensor and noise tensor : (1) in , representing the tensors of the original image, background, target, and noise, respectively; S2. The background tensor for estimating small infrared targets contains a large amount of noise caused by clouds, sea clutter, etc., resulting in highly redundant spatial information. This redundant spatial information leads to severe degradation of the original infrared background image. To mimic the real background degradation process, we not only follow existing work (directly modeling the background tensor) but also consider a linear process. The degradation model obtained from step S1 is as follows: A linear process is used to simulate the real background degradation process. (2); S3. Construct a low-rank learning model for structured background tensor estimation, and then use the background tensor... Projected onto a set of bases In the resulting subspace, the degenerate model is transformed into a regularized tensor estimation model: background tensor Decompose into base and coefficient tensor , According to the maximum a posteriori estimation theory, the degenerate tensor model can be transformed into a regularized tensor estimation model: (3) in It is a coefficient tensor Constraints, It is a regularization parameter. Represents the tensor modulus product, The Frobenius norm is a regularization term used in image restoration problems. S4. Due to factors such as imaging distance and environment, small infrared targets lack texture and spatial structure information, occupying only a small portion of the image. Therefore, the target can be considered as a sparse matrix; while many local patches in the background image are approximately linearly correlated, such as... Figure 2 Therefore, the background tensor This can be viewed as a low-rank matrix. Therefore, the infrared small target detection problem is transformed into: (4) in These are compromise parameters used to control the target tensor and the background tensor. This represents the target tensor. Norm, Represents the nuclear norm; S5, Coefficient Tensor Obtained from Background Tensor Decomposition The coefficient tensor contains redundant spatial information, therefore the coefficient tensor... Regularization constraints are applied, and tensor structured cascaded rank decomposition is used to effectively utilize... The inherent correlation and low rank of the core tensor By decomposing and inverting in a cascaded manner, the spatial redundancy information of the degenerate tensor is removed, thus the coefficient tensor... constraint terms It can be represented as: (5) in It is a linear transformation operator. Represents the coefficient tensor slices, Represents the core tensor. Indicates to The proposed tensor cascade rank minimization; S6, Subspace and coefficient tensor The constraint terms are added to the background tensor model to obtain: (6)

[0025] Further, this can be divided into the following two sub-problems:

[0026] By introducing the Alternating Direction Multiplier Method (ADMM) to solve the model, the augmented Lagrangian function can be defined as: ,

[0027] in It is a Lagrange multiplier. It is a penalty factor greater than 0. This represents the inner product of two third-order tensors; subsequently, the alternating direction multiplier method is used to... Minimize it by decomposing it into several subproblems, and It can be updated using an iterative method.

[0028] S7. Optimize the solution of subproblems.

[0029] for :

[0030] for :

[0031] for :

[0032] for :

[0033] Regarding the introduced parameters:

[0034] in As a control factor greater than 1, in the specific experiments of this invention, .

[0035] The real target image and its 3D image after filtering the above-mentioned infrared small target using the method of the present invention are shown below. Figure 3 As shown, the improved target saliency is clearly evident. The detection method of this invention effectively suppresses interference from complex backgrounds, achieving accurate detection of small targets. The following comparative experiment further verifies the effectiveness of this invention. Figure 2 The original infrared images were tested using the IPI, RLCM, SRWS, and NOLC tests, and the results are as follows: Figure 4-7 As shown, and with Figure 3 The experimental results were compared to those of other methods, further verifying the significant advantages of the infrared small target detection method provided by this invention. The above are merely preferred embodiments of this invention and do not constitute any limitation on the invention. Any equivalent substitutions or modifications made by those skilled in the art to the technical solutions and content disclosed in this invention without departing from the scope of the invention's technical solutions are considered to have remained within the protection scope of this invention.

Claims

1. An infrared small target detection method based on structured background tensor estimation, characterized in that, Includes the following steps: S1. Divide the original image tensor into target tensors. Background tensor and noise tensor ; S2. Model the background tensor using a linear process. A degradation model is obtained by simulating the real background degradation process; S3. Construct a low-rank learning model for structured background tensor estimation, projecting the background tensor onto a set of base values. In the resulting subspace, the degenerate model is transformed into a regularized tensor estimation model; S4. Coefficient tensor obtained from background tensor decomposition Regularization constraints are applied to decompose and invert the core tensor in a cascaded manner to remove the spatial redundancy information of the degenerate tensor. S5, Subspace and coefficient tensor The constraint terms are added to the structured background tensor estimation model, and the formula is optimized using the alternating direction multiplier method. S6. Optimize and solve for each parameter, train the entire model, and realize infrared small target detection; In step S3, the background tensor is... Decompose into base and coefficient tensor , According to the maximum a posteriori estimation theory, the degenerate tensor model is transformed into a regularized tensor estimation model: , in, It is a coefficient tensor Constraints, It is a regularization parameter. Represents the tensor modulus product, The Frobenius norm is a regularization term used in image restoration problems. The target is set as a sparse matrix, and the background tensor is... Setting it as a low-rank matrix, the infrared small target detection problem is transformed into: ; In step S4, the coefficient tensor obtained from the background tensor decomposition For coefficient tensors Apply regularization constraints to the core tensor The decomposition and inversion are performed in a cascaded manner to remove spatial redundancy information from the degenerate tensor and the coefficient tensor. constraint terms Represented as: , in It is a linear transformation operator. Represents the coefficient tensor slices, Represents the core tensor. Indicates to The proposed tensor cascade rank minimization; Step S5 will adjust the subspace and coefficient tensor Adding constraint terms to the background tensor model yields: .

2. The infrared small target detection method based on structured background tensor estimation according to claim 1, characterized in that, In step S1, the original image tensor is described as follows: .

3. The infrared small target detection method based on structured background tensor estimation according to claim 1, characterized in that, In step S2, a linear process is used to simulate the real background degradation process, and the resulting degradation model is as follows: .

4. The infrared small target detection method based on structured background tensor estimation according to claim 3, characterized in that, The Alternating Direction Multiplier Method (ADMM) algorithm is used to solve the model: 。

Images