A thermal infrared small target detection method based on non-overlapping block space-time tensor model

Abstract

The application discloses a thermal infrared small target detection method based on a non-overlapping block space-time tensor model. The method comprises the following steps: (1) using an original thermal infrared image sequence to construct a non-overlapping block space-time tensor according to a specific rule; (2) converting the infrared weak small target detection task into a tensor robust principal component analysis problem, and building a low-rank sparse tensor decomposition framework; (3) using the multi-mode expansion of the tensor and a Laplace function to obtain a non-convex low-rank estimation norm of a background tensor; (4) using a re-weighting strategy to obtain a sparsity estimation of a target tensor; (5) using a tubular sparse regularization term to measure a sparse structure component, and using a Frobenius norm to measure noise; and (6) optimizing and solving the model based on an improved ADMM algorithm, obtaining a target tensor component, reconstructing a target detection result image, and realizing infrared weak small target detection. The low-rank sparse tensor decomposition framework and the improved ADMM optimization algorithm can effectively realize the detection of infrared weak small targets.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to a thermal infrared small target detection method based on a non-overlapping block spatiotemporal tensor model. Background Art

[0002] In recent years, infrared search and tracking systems have found widespread application in military and military fields such as mine detection, low-altitude security, and missile tracking. Robust infrared small target detection plays a crucial role in these systems. However, due to long-range imaging, the size of small targets in infrared images typically ranges from 2×2 to 9×9 pixels, and they often lack specific shape, texture details, and other structural information. Furthermore, due to the complex imaging environment, the background of infrared images is intricate, and small targets with low signal-to-noise ratios are often obscured by various interferences, such as dense clouds and ocean clutter. In such complex backgrounds, it is difficult to simultaneously improve the algorithm's target detectability and background suppression capabilities. Consequently, infrared small target detection still faces numerous challenges.

[0003] Currently, scholars at home and abroad have conducted extensive research in the field of infrared small target detection, proposing various methods for detecting weak infrared targets, including model-driven and deep learning-based methods. Single-frame detection methods can be roughly divided into three categories: background suppression-based, human visual system (HVS)-based, and low-rank sparse representation-based. Background suppression-based methods typically design filters to estimate the background, filtering it out from the original image to detect the target, but their detection performance is relatively poor in complex noisy scenes. HVS-based methods assume that humans can distinguish small targets based on the brightness difference between the target and the background. The most classic method is the local contrast method (LCM), but the detector's detection ability is easily affected by the background. Low-rank sparse representation-based methods can be divided into matrix-based and tensor-based methods. They achieve target component separation through low-rank background estimation, sparse target estimation, and random noise measurement. Image patches are obtained by sliding a window over the infrared image and reconstructed into a new image matrix. Low-rank sparse matrix decomposition is performed. However, the process of constructing the matrix significantly destroys the spatial structure of the original infrared image, which to some extent limits the performance of infrared small target detection. Therefore, we expand the matrix domain to the tensor domain, constructing a tensor from image blocks, and utilizing the correlation between different blocks to perform a low-rank sparse decomposition of the tensor. Since most existing methods fail to accurately estimate the background low-rank and target sparsity, while low-rank sparse representations achieve good results in simple backgrounds, they have certain limitations in complex backgrounds.

[0004] With the rapid development of deep learning applications in various fields, infrared small target detection based on deep learning has garnered widespread attention. Neural networks possess powerful feature learning capabilities. To overcome the shortcomings of traditional methods, infrared small target detection can be modeled as a supervised machine learning problem, employing neural networks as both feature extractors and detectors. For example, ACM and AlcNet fuse low-level and high-level features to highlight and preserve small target features. To leverage convolutional neural networks, MDvsFA-cGAN uses a shallow CNN to achieve high-accuracy infrared small target detection and a relatively deep CNN to detect targets with a low false negative rate. The final detection result is obtained by fusing the results from the shallow and deep CNNs. However, supervised infrared small target detection or segmentation remains challenging for many existing baseline networks. Many deep networks can fail due to the scarcity of intrinsic features of small targets and the presence of background clutter. Furthermore, most networks learn high-level semantic features by gradually decaying the size of feature maps, making it easy for small targets to be overwhelmed by surrounding background features in the deeper layers. Currently, no infrared small target detection algorithm achieves good detection results under complex backgrounds and low signal-to-noise ratio conditions. Summary of the Invention

[0005] In view of the deficiencies in the prior art, the purpose of the present invention is to provide a thermal infrared small target detection method based on a non-overlapping block spatiotemporal tensor model, which can realize the detection of infrared weak small targets in complex backgrounds, while improving the detection rate of thermal infrared small targets and reducing the false detection rate.

[0006] Leveraging the low-rank nature of the background and the sparse nature of the target in infrared images, the original infrared image sequence is converted into a non-overlapping spatiotemporal tensor. The impact of background, noise, and rare structural components on infrared small target detection is comprehensively considered, transforming the small target detection problem into a low-rank sparse tensor decomposition optimization problem. A framework for infrared small target detection is then constructed. While preserving the tensor structure, the spatiotemporal structure of the background tensor in different dimensions is utilized to more realistically represent its low-rank properties. An improved ADMM algorithm is used to efficiently solve the optimization model, enabling infrared small target detection.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] The present invention provides a thermal infrared small target detection method based on a non-overlapping block spatiotemporal tensor model, which is characterized by comprising the following steps:

[0009] Step 1): Using the original thermal infrared image sequence D, a fixed-size window is slid on each frame image with a fixed sliding step, and the obtained image blocks are stacked in sequence to realize the spatiotemporal tensor from the original image sequence to non-overlapping blocks. transformation;

[0010] Step 2): Convert the thermal infrared small target detection task into a tensor robust principal component analysis problem and build a low-rank sparse tensor decomposition framework;

[0011] Step 3): Using the mode-k1k2 expansion of the tensor and the sum of the Laplace function values ​​of the singular values, a non-convex low-rank estimation norm is constructed, namely the multi-mode weighted tensor kernel norm ||·|| MWTNN , the metric background tensor Low-rank properties;

[0012] Step 4): Estimate the target tensor The sparsity of the algorithm is calculated and the sparsity reweighting strategy is used to speed up the algorithm solution.

[0013] Step 5): Use l 1,1,2 The norm is used as a tubular sparse regularization term to measure the linear sparse structure tensor ε of strong edges, and the Frobenius norm is used to measure the noise tensor

[0014] Step 6): After steps 3) to 5), the low-rank sparse tensor decomposition framework is transformed into a solvable optimization model, and the optimization model is solved to obtain the target tensor components. Then, according to the inverse transformation method of constructing the non-overlapping block spatiotemporal tensor in step 1), the target detection result image sequence T is reconstructed to obtain the thermal infrared weak target detection result.

[0015] In summary, the technical solution conceived by the present invention has the following advantages compared with the prior art:

[0016] (1) The present invention comprehensively considers the influence mechanism of background, noise and linear sparse structure on infrared small target detection, and builds an infrared small target detection framework, which essentially transforms the target detection problem into a low-rank sparse tensor decomposition optimization problem, which can alleviate the contradiction between the detector's target detection performance and background suppression performance to a certain extent.

[0017] (2) The present invention proposes a new non-convex low-rank estimation norm, namely the multi-modal weighted tensor kernel norm, which fully considers the spatiotemporal structural information of the background tensor in different dimensions, can more realistically characterize the low-rank characteristics of the background tensor, and is conducive to background separation.

[0018] (3) This paper proposes a new non-overlapping block spatiotemporal tensor construction method, and integrates the proposed tensor non-convex low-rank estimation and tubular sparse regularization term into the non-overlapping spatiotemporal tensor decomposition model. On this basis, an improved ADMM algorithm is proposed to efficiently solve the optimization model. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1This is an overall flow chart of an embodiment of the thermal infrared image small target detection method of the present invention.

[0020] Figure 2 Schematic diagram of constructing non-overlapping block tensors for a single-frame thermal infrared image in the present invention.

[0021] Figure 3 Schematic diagram of constructing non-overlapping block space-time tensors for thermal infrared image sequences in the present invention.

[0022] Figure 4 Example frame images of thermal infrared image sequences used for experimental testing.

[0023] Figure 5 Comparison of small target detection results after thermal infrared images are detected using different methods and the corresponding three-dimensional grayscale visualization images. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention is described in detail below with reference to specific embodiments. Specific embodiments are described below to simplify the present invention. However, it should be understood that the present invention is not limited to the illustrated embodiments, and that various modifications of the present invention are possible without departing from the underlying principles, and these equivalent forms also fall within the scope defined by the appended claims.

[0025] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0026] The embodiment of the present invention is a thermal infrared small target detection method based on a non-overlapping block spatiotemporal tensor model. The specific implementation process is as follows: Figure 1 Shown, including:

[0027] Step 1): Using the original thermal infrared image sequence D, a fixed-size window is slid on each frame image with a fixed sliding step, and the obtained image blocks are stacked in sequence to realize the spatiotemporal tensor from the original image sequence to non-overlapping blocks. transformation;

[0028] Step 2): Convert the thermal infrared small target detection task into a tensor robust principal component analysis problem and build a low-rank sparse tensor decomposition framework;

[0029] Step 3): Using the mode-k1k2 expansion of the tensor and the sum of the Laplace function values ​​of the singular values, a non-convex low-rank estimation norm is constructed, namely the multi-mode weighted tensor kernel norm ||·|| MWTNN , the metric background tensor Low-rank properties;

[0030] Step 4): Estimate the target tensor The sparsity of the algorithm is calculated and the sparsity reweighting strategy is used to speed up the algorithm solution.

[0031] Step 5): Use l 1,1,2 The norm is used as a tubular sparse regularization term to measure the linear sparse structure tensor ε of strong edges, and the Frobenius norm is used to measure the noise tensor

[0032] Step 6): After steps 3) to 5), the low-rank sparse tensor decomposition framework is transformed into a solvable optimization model, and the optimization model is solved to obtain the target tensor components. Then, according to the inverse transformation method of constructing the non-overlapping block spatiotemporal tensor in step 1), the target detection result image sequence T is reconstructed to obtain the thermal infrared weak target detection result.

[0033] In a specific embodiment of the present invention, the step 1) is specifically as follows:

[0034] Using original thermal infrared image sequences The obtained n(=n r ×n c ) image blocks are stacked as front slices to obtain non-overlapping image block tensors like Figure 2 As shown in the figure, the tensors corresponding to the adjacent k frames are stacked in sequence to obtain non-overlapping block spatiotemporal tensors rich in temporal and spatial information. like Figure 3 As shown in the example, the value of k is 3 and the value of ps is 15;

[0035] Furthermore, the step 2) is specifically as follows: non-overlapping block spatiotemporal tensor Can be regarded as a background tensor Target tensor and the noise tensor The linear combination of

[0036]

[0037] Considering the low-rank characteristics of the background and the sparse characteristics of the target, small target detection can be transformed into a tensor robust principal component analysis problem, and a low-rank sparse tensor decomposition framework is constructed, as shown in formula (2):

[0038]

[0039] Considering that non-target objects with streamline appearance in complex situations usually exhibit the property of linear sparse structures, and anomalies generated by these structures (e.g., prominent building edges) may be misdetected as target components, the non-overlapping block spatiotemporal tensor is transformed into Modeled as a background tensor Target tensor Sparse structured component tensor ε and noise tensor The linear combination of is calculated as:

[0040]

[0041] In a specific embodiment of the present invention, the step 3) is specifically as follows:

[0042] Non-overlapping block space-time tensors The first two dimensions of have rich spatial information, which together describe the local correlation of each patch. The third dimension contains sufficient temporal and spatial features, from which the non-local features of the background can be mined. In order to preserve the structural information between multiple modes of the tensor and improve the computational efficiency, the background tensor Perform the mode-k1k2 expansion, and the calculation formula is:

[0043]

[0044] in, Represents the pattern-k1k2 spread operator for a tensor.

[0045] Since different singular values ​​in infrared images have clear physical meanings, they need to be given different weights to prevent image information loss. All frontal slices B (k) The sum of the Laplace function values ​​of the singular values ​​of (k=1,2,…,n3) is used as the improved tensor nuclear norm to characterize the low-rank characteristics of the background tensor. The calculation formula is:

[0046]

[0047] Among them, the tensor The dimensions are n1×n2×n3, k=1,2,…,n3, is the background tensor The kth frontal slice after fast Fourier transform The jth singular value of g(·) is the Laplace function, defined as x is the independent variable, ε is a positive constant;

[0048] Preferably, the non-convex low-rank estimation norm (multi-mode weighted tensor kernel norm||·||) that can deeply mine the spatiotemporal information of the background tensor and characterize its low-rank characteristicsMWTNN ), the calculation formula is:

[0049]

[0050] in, is the weighting coefficient used to weight the background tensor Improved tensor nuclear norm of all tensors obtained after pattern-k1k2 expansion is a tensor The i-th frontal slice of .

[0051] In a preferred embodiment of the present invention, in step 4), the l1 norm is used instead of the l0 norm in step 2), so that the sparsity measure of the target tensor is solvable, and a sparse reweighting strategy is used to speed up the algorithm solution. The iterative calculation formula of the sparsity enhancement weight is:

[0052]

[0053] Among them, a is a positive number, b is a positive number to prevent the denominator from being zero, j is the number of iterations, is the target tensor in the jth iteration The absolute value of The sparsity enhancement weight of the j+1th generation can be obtained Specifically in the embodiment, the value of a is 2, and the value of b is 0.01.

[0054] Furthermore, in the step 5), the method of using 1,1,2 The norm is used as a tubular sparsity regularization term to measure the sparse linear structure tensor ε, and the calculation formula is:

[0055]

[0056] Among them, ||·|| 1,1,2 l 1,1,2 norm, ||·||2 is the l2 norm, and ε(i,l,j) is the element in the i-th row and l-th column of the j-th frontal slice in the sparse linear structure tensor ε.

[0057] In a preferred embodiment of the present invention, the specific implementation of step 6) is as follows:

[0058] The low-rank sparse tensor decomposition framework (3) is transformed into a solvable optimization model as shown in Equation (9).

[0059]

[0060] Among them, ||·|| MWTNNis the multi-mode weighted tensor kernel norm, λ1, λ2 and λ3 are coordination parameters, ||·||1 is the l1 norm, ||·|| F represents the Frobenius norm, ⊙ is the Hadamard product, Enhance weights for sparsity.

[0061] Preferably, the optimization model is solved based on the improved ADMM algorithm. Since the constructed non-overlapping block space-time tensor is a three-dimensional tensor, the background tensor can be expanded through mode-12, mode-13, and mode-23 to obtain three tensors respectively. Right now Therefore, by introducing three auxiliary variables Right now The equivalent model is shown in formula (10):

[0062]

[0063] Among them, ||·|| MWTNN is the multi-mode weighted tensor kernel norm, λ1, λ2 and λ3 represent coordination parameters, ||·||1 represents the l1 norm, ||·|| F represents the Frobenius norm, 1≤k1<k2≤3;

[0064] The calculation formula of the augmented Lagrangian function of formula (10) is:

[0065]

[0066] Among them, ||·|| MWTNN is the multi-mode weighted tensor nuclear norm, ρ and is the penalty parameter, and is the Lagrange multiplier, is the independent variable, and the three auxiliary variables The improved tensor nuclear norm weight coefficient is α 12 =ω / (2+ω),α 13 =α 23 =1 / (2+ω), ω is a custom parameter, is the square of the Frobenius norm; specifically in the embodiment, the value of ω is 0.001.

[0067] Since it is difficult to solve all the variables in (10) at the same time, (10) is decomposed into several sub-problems and the variables are updated alternately.

[0068] -Subproblem: Fix other parameters of generation j and iteratively update generation j+1 The calculation formula is

[0069]

[0070] in, Representing a tensor Operators restored according to the pattern -k1k2, represents the singular value contraction operator, σ is the singular value, τ is the threshold, diag[·] is the operator for taking the diagonal elements of the matrix, ζ + =max(ζ,0) means taking the maximum value, For tensors Perform t-SVD decomposition, “*” is the tensor product operation, thus we can get and represents the Laplace function The derivative of is a tensor The sth singular value of the ith frontal slice, σ (i)1 ≥σ (i)2 ≥…≥σ (i)s Auxiliary variables The i-th frontal slice The singular values ​​of

[0071] -Subproblem: Fix other parameters of generation j and iteratively update generation j+1 The calculation formula is

[0072]

[0073] -Subproblem: Fix other parameters of generation j and iteratively update generation j+1 The calculation formula is

[0074]

[0075] in, is the soft threshold shrinkage operator, sign(·) is the sign function, ξ is the soft threshold;

[0076] ε-subproblem: fix other parameters of the jth generation and iteratively update the j+1th generation ε (j+1) The calculation formula is

[0077]

[0078]

[0079] -Subproblem: Fix other parameters of generation j and iteratively update generation j+1 The calculation formula is

[0080]

[0081] ρ, -Sub-questions:

[0082]

[0083]

[0084] ρ (j+1) =min(κρ (j) ,ρ max ) (20)

[0085]

[0086] Among them, κ is the amplification coefficient, which makes the penalty parameter ρ of the j+1 generation, is κ times of the jth generation, ρ max and is the maximum value of the penalty parameter.

[0087] During the iteration process, when the relative error When it is less than δ, the iteration is terminated; in order to improve the computational efficiency, when the target tensor The number of non-zero elements in ) no longer changes, and the iteration stops. Specifically in the example, λ2=λ3=50λ1, λ L The value of is 1.4, The value of is 5, the value of κ is 1.5, and the value of δ is 10 -4 For the final target tensor component According to the inverse transformation method of constructing non-overlapping block spatiotemporal tensors as described in claim 2, the target detection result image sequence T is reconstructed to realize thermal infrared dim small target detection.

[0088] Compared with the prior art, the method of the embodiment of the present application comprehensively considers the influence of background, noise and rare structure components on infrared small target detection, and builds an infrared small target detection framework, which essentially converts the target detection problem into a low-rank sparse tensor decomposition optimization problem, and can relieve the contradiction between the target detection performance and the background suppression performance of the detector; a new non-convex low-rank estimation norm, i.e., a multi-mode weighted tensor nuclear norm, is proposed, which fully considers the spatiotemporal structure information of the background tensor in different dimensions, can more truly represent the low-rank characteristics of the background tensor, and is beneficial to realize background separation; a non-overlapping block spatiotemporal tensor construction method is proposed, and the proposed tensor non-convex low-rank estimation and tubular sparse regularization term are integrated into a non-overlapping spatiotemporal tensor decomposition model, and on this basis, an improved ADMM algorithm is proposed to efficiently solve the optimization model.

[0089] Embodiment

[0090] Next, the disclosed infrared image sequence is taken as the research object, and the thermal infrared small target detection algorithm is verified. In order to evaluate the small target detection result from the qualitative and quantitative angles, the small target detection result image and 3D-ROC and its derivative evaluation indexes including AUC TD , AUC BS , AUC SNPR , AUC TDBS and AUC ODP are respectively used to comprehensively evaluate the small target detection result from the angles of target detection capability, background suppression capability and detector effectiveness. The AUC (D,F) , AUC (D,τ) , AUC (F,τ) included in 3D-ROC can respectively represent the detector effectiveness, target detection capability and background suppression capability, and the derivative evaluation indexes thereof are also divided according to the corresponding evaluation angles, and the specific definitions are as follows:

[0091] AUC (D,F) represents the AUC value of the (P D , P F ) curve, representing the effectiveness of the detector,

[0092] AUC (D,F) ∈[0,1]

[0093] AUC (D,τ) represents the AUC value of the (P D , τ) curve, representing the target detection capability of the detector,

[0094] AUC (D,τ) ∈[0,1]

[0095] AUC (F,τ) represents the AUC value of the (P D,τ) curve, characterizes the background suppression ability of the detector.

[0096] AUC (F,τ) ∈[0,1]

[0097] AUC TD represents the joint detection capability of the detector,

[0098] AUC TD =AUC (D,F) +AUC (D,τ) )∈[0,2]

[0099] AUC BS represents the joint background suppression capability of the detector,

[0100] AUC BS =AUC (D,F) -AUC (F,τ) )∈[-1,1]

[0101] AUC TDBS represents the detector’s comprehensive ability of target detection and background suppression,

[0102] AUC TDBS =AUC (D,τ) -AUC (F,τ) )∈[-1,1]

[0103] AUC SNPR represents the signal-to-noise ratio of the detector,

[0104]

[0105] AUC ODP Represents the total detection probability of the detector

[0106] AUC ODP =AUC (D,F) +AUC (D,τ) -AUC (F,τ) ∈[-1,2]

[0107] In general, the evaluation indicators are divided as follows:

[0108] (i) Target Detection (TD): AUC (D,τ) ,AUC TD

[0109] (ii) Background suppression (BS): AUC (F,τ) ,AUC BS ,AUC SNPR

[0110] (iii) Detector effectiveness: AUC (D,F),AUC TDBS ,AUC ODP

[0111] For a public real thermal infrared image sequence of 100 frames, its characteristics are: sky-ground background, a small helicopter with low local contrast, a suspected target in the background, thick clouds and severe noise, and the image size is 256×256. The example frame of the thermal infrared image sequence is as follows: Figure 4 Table 1 shows the quantitative indicators of small target detection results using IPI, NRAM, RIPT, WSNM-STIPT, TCNN-NPSTT, ASTTV-NTLA and the method of the present invention in thermal infrared image sequences.

[0112] Table 1 Quantitative indicators of the detection results of thermal infrared images using IPI, NRAM, RIPT, WSNM-STIPT, TCNN-NPSTT, ASTTV-NTLA and the method of the present invention

[0113] method <![CDATA[AUC (D,F) ]]> AUC (D,τ) ]] <![CDATA[AUC (F,τ) ]]> <![CDATA[AUC TD ]]> <![CDATA[AUC BS ]]> <![CDATA[AUC SNPR ]]> <![CDATA[AUC TDBS ]]> AUC ODP ]]> The present invention 1.0000 1.0000 0.0050 2.0000 0.9950 2.0000e2 0.9950 1.9950 IPI 1.0000 1.0000 0.1903 2.0000 0.8097 5.2538 0.8097 1.8097 NRAM 1.0000 0.9903 0.0051 ]]> ​ 1.9902 <![CDATA[ 0.9949 ]]> <![CDATA[ 1.9436e2 ]]> <![CDATA[ 0.9852 ]]> <![CDATA[ 1.9852 ]]> RIPT 1.0000 0.9745 0.0053 1.9745 0.9947 1.8376e2 0.9692 1.9692 WSNM-STIPT 1.0000 0.9634 <![CDATA[ 0.0051 ]]> 1.9633 0.9949 ]]> ​ 1.8980e2 0.9583 1.9583 TCNN-NPSTT 1.0000 0.9649 <![CDATA[ 0.0051 ]]> 1.9648 <![CDATA[ 0.9949 ]]> 1.9067e2 0.9598 1.9598 ASTTV-NTLA 1.0000 <![CDATA[ 0.9942 ]]> 0.0589 <![CDATA[ 1.9941 ]]> 0.9411 1.6891e1 0.9353 1.9353

[0114] Figure 5 The comparison diagram of small target detection results and the corresponding three-dimensional grayscale visualization images for thermal infrared image sequences using IPI, NRAM, RIPT, WSNM-STIPT, TCNN-NPSTT, ASTTV-NTLA and the method of the present invention are combined with the quantitative detection results of the eight evaluation indicators shown in Table 1. It can be seen that the method proposed in the present invention is superior to the comparison method in all eight 3D-ROC derived evaluation indicators. From a general point of view, whether it is background suppression or target enhancement effect, the proposed method is superior to the comparison algorithm. In summary, the superiority of the thermal infrared small target detection method proposed in the present invention in target enhancement and background suppression can be reflected at both the visual effect and quantitative indicators levels, and it has excellent thermal infrared small target detection capabilities.

[0115] The accompanying drawings illustrating the embodiments of the present invention serve to more clearly illustrate the objectives, technical solutions, and advantages of the present invention. It should be noted that the specific embodiments described herein are intended only to illustrate the present invention and are not intended to limit the present invention. Any equivalent substitutions, modifications, and the like made within the methodologies and principles provided by the present invention are intended to be included within the scope of protection of the present invention.

Claims

1. A thermal infrared small target detection method based on a non-overlapping block spatiotemporal tensor model, characterized in that: The steps include: Step 1): Using the original thermal infrared image sequence D, a fixed-size window is slid on each frame image with a fixed sliding step, and the obtained image blocks are stacked in sequence to realize the spatiotemporal tensor from the original image sequence to non-overlapping blocks. transformation; Step 2): Convert the thermal infrared small target detection task into a tensor robust principal component analysis problem and build a low-rank sparse tensor decomposition framework; Step 3): Using the mode-k1k2 expansion of the tensor and the sum of the Laplace function values ​​of the singular values, a non-convex low-rank estimation norm is constructed, namely the multi-mode weighted tensor kernel norm ||·|| MWTNN , the metric background tensor Low-rank properties; Step 4): Estimate the target tensor The sparsity of the algorithm is calculated and the sparsity reweighting strategy is used to speed up the algorithm solution. Step 5): Utilize The norm is used as a tubular sparse regularization term to measure the linear sparse structure tensor ε of strong edges, and the Frobenius norm is used to measure the noise tensor Step 6): After steps 3) to 5), the low-rank sparse tensor decomposition framework is transformed into a solvable optimization model, and the optimization model is solved to obtain the target tensor components. Then, according to the inverse transformation method of the non-overlapping block spatiotemporal tensor constructed in step 1), the target detection result image sequence T is reconstructed to obtain the thermal infrared weak target detection result; The step 2) is specifically as follows: Non-overlapping block space-time tensors Considered as background tensor Target tensor and the noise tensor The linear combination of is: Taking into account the low-rank characteristics of the background and the sparse characteristics of the target, the thermal infrared small target detection task is transformed into a tensor robust principal component analysis problem, and a low-rank sparse tensor decomposition framework is constructed, as shown in formula (2): Among them, rank(·) represents the rank calculation operator, ||·||0 represents norm, ||·|| F represents the Frobenius norm, λ1 represents the coordination parameter, and δ represents the noise error term; Considering that non-target objects with streamline appearance in complex situations usually exhibit the property of linear sparse structures, and the anomalies generated by these structures may be misdetected as target components, the non-overlapping block space-time tensor is converted into Modeled as a background tensor , target tensor , sparse structure tensor ε and noise tensor The linear combination of is:

2. The thermal infrared small target detection method based on the non-overlapping block spatiotemporal tensor model according to claim 1 is characterized in that: The step 1) is specifically as follows: Using the original thermal infrared image sequence f, a fixed-size sliding window is slid in an S-shaped path on each frame of the image with a certain sliding step size. The obtained non-overlapping image blocks are stacked in sequence as frontal slices to obtain the non-overlapping image block tensor of each frame of the image; the non-overlapping image block tensors corresponding to the adjacent k frames of the image are stacked in sequence to obtain the non-overlapping block spatiotemporal tensor rich in temporal and spatial information. .

3. The thermal infrared small target detection method based on the non-overlapping block spatiotemporal tensor model according to claim 1 is characterized in that: The step 3) is specifically as follows: Background tensor Perform the mode-k1k2 expansion, and the calculation formula is: in, The k1k2 expansion operator represents the pattern of a tensor. Represents the background tensor The three-dimensional tensor after expansion by mode -k1k2, Indicates that the dimension of the three-dimensional tensor is Π represents the product operator; The background tensor All frontal slices after fast Fourier transform The sum of the Laplace function values ​​of the singular values ​​is used as the improved tensor nuclear norm Characterizing the low-rank characteristics of the background tensor, the calculation formula is: Among them, the tensor The dimensions are n1×n2×n3, k=1,2,…,n3, is the background tensor The kth frontal slice after fast Fourier transform The jth singular value of g(·) is the Laplace function, defined as x is the independent variable, ε is a positive constant; The multi-mode weighted tensor nuclear norm ||·|| MWTNN , the calculation formula is: in, is the weighting coefficient used to weight the background tensor Improved tensor nuclear norm of all tensors obtained after pattern-k1k2 expansion is a tensor The i-th frontal slice of .

4. The thermal infrared small target detection method based on non-overlapping block spatiotemporal tensor model according to claim 1 is characterized in that: In the step 4) described above, In the norm optimal approximation step 2) norm, making the sparsity metric of the target tensor solvable, and using the sparse reweighting strategy to speed up the algorithm solution. The iterative calculation formula of the sparsity enhancement weight is: Among them, a is a positive number, b is a positive number to prevent the denominator from being zero, j is the number of iterations, is the target tensor in the jth iteration The absolute value of The sparsity enhancement weight of the j+1th generation can be obtained 5. The thermal infrared small target detection method based on non-overlapping block spatiotemporal tensor model according to claim 1 is characterized in that: In the step 5) described above, The norm is used as a tubular sparsity regularization term to measure the sparse linear structure tensor ε, and the calculation formula is: Among them, ||·|| 1,1,2 for norm, ||·||2 is norm, ε(i,l,j) is the element in the i-th row and l-th column of the j-th frontal slice in the sparse linear structure tensor ε.

6. The thermal infrared small target detection method based on non-overlapping block spatiotemporal tensor model according to claim 1 is characterized in that: The specific implementation of step 6) is as follows: The low-rank sparse tensor decomposition framework is transformed into a solvable optimization model, as shown in formula (9), Among them, ||·|| MWTNN is the multi-mode weighted tensor nuclear norm, λ1, λ2 and λ3 are coordination parameters, and ||·||1 is norm, ||·|| F represents the Frobenius norm, ⊙ is the Hadamard product, Enhance weights for sparsity; The optimization model is solved based on the improved ADMM algorithm. Since the constructed non-overlapping block space-time tensor is a three-dimensional tensor, the background tensor can be expanded through mode-12, mode-13, and mode-23 to obtain three tensors respectively. Right now Therefore, by introducing three auxiliary variables Right now The equivalent model is shown in formula (10): Among them, ||·|| MWTNN is the multi-mode weighted tensor nuclear norm, λ1, λ2 and λ3 represent coordination parameters, and ||·||1 represents norm, ||·|| F represents the Frobenius norm, 1≤k1<k2≤3; The augmented Lagrangian function of formula (10) The calculation formula is: Among them, ||·|| MWTNN is the multi-mode weighted tensor nuclear norm, ρ and is the penalty parameter, and is the Lagrange multiplier, is the independent variable, and the three auxiliary variables The improved tensor nuclear norm weight coefficient is α 12 =ω / (2+ω),α 13 =α 23 =1 / (2+ω), ω is a custom parameter, is the square of the Frobenius norm; Decompose formula (10) into several sub-problems and update the variables alternately: -Subproblem: Fix other parameters of generation j and iteratively update generation j+1 The calculation formula is Among them, t-fold( ,k1,k2) represents a tensor Operators restored according to the pattern -k1k2, represents the singular value contraction operator, σ is the singular value, τ is the threshold, diag[·] is the operator for taking the diagonal elements of the matrix, ζ + =max(ζ,0) means taking the maximum value, For tensors Perform t-SVD decomposition, "*" is the tensor product operation, thus we can get and represents the Laplace function The derivative of is a tensor The sth singular value of the ith frontal slice, σ (i)1 ≥σ (i)2 ≥…≥σ (i)s Auxiliary variables The i-th frontal slice The singular values ​​of -Subproblem: Fix other parameters of generation j and iteratively update generation j+1 The calculation formula is -Subproblem: Fix other parameters of generation j and iteratively update generation j+1 The calculation formula is in, is the soft threshold shrinkage operator, sign(·) is the sign function, ξ is the soft threshold; ε-subproblem: fix other parameters of the jth generation and iteratively update the j+1th generation ε (j+1) The calculation formula is -Subproblem: Fix other parameters of generation j and iteratively update generation j+1 The calculation formula is ρ, -Sub-questions: r (j+1) =min(kr (j) ,r max ) (20) Among them, κ is the amplification coefficient, which makes the penalty parameter ρ of the j+1 generation, is κ times of the jth generation, ρ max and is the maximum value of the penalty parameter set; During the iteration process, when the relative error When it is less than η, the iteration terminates; when the target tensor The number of nonzero elements in No more changes, no more iterations; Finally get the target tensor component According to the inverse transformation method of constructing non-overlapping block spatiotemporal tensors, the target detection result image sequence T is reconstructed to obtain the thermal infrared weak target detection result.

Images