A small target detection method based on decomposition model

Abstract

The present invention provides a small target detection method based on a decomposition model, which belongs to the technical field of infrared small target detection and solves the real-time and accuracy defects existing in the existing technical methods. The method comprises: S1, using a preliminary decomposition model to process infrared images, and determining potential targets by constructing a tensor model; S2, based on an entropy matching method, performing motion monitoring on potential targets, and calculating motion change information between image frames corresponding to the potential targets; S3, for image frames whose motion change information reaches a preset threshold, calculating the weight tensor that needs to be completely decomposed in this image frame, and activating the complete decomposition model; S4, using the complete decomposition model to process the weight tensor to obtain a decomposition result. The present invention no longer uses a block tensor method, but directly decomposes the image frame. By combining the preliminary decomposition and the complete decomposition methods, the detection result is ultimately made to have higher accuracy and lower false alarm rate.

Description

Technical Field

[0001] The present invention belongs to the technical field of infrared small target detection, applies the concept of image frame decomposition, and is specifically a small target detection method based on a decomposition model. Background Art

[0002] In the research of small target detection in infrared image sequences, related detection methods always face unchangeable problems: 1. Small targets have dim brightness, very few pixels, and lack texture and shape information; 2. The contrast between small targets and image background is very low, and they are greatly affected by noise, clutter and illumination changes; 3. Existing detection algorithms cannot achieve both high real-time performance and robustness while ensuring detection accuracy.

[0003] With the advancement of related research, the detection of small infrared targets using low-rank sparse decomposition models has become a very effective detection method in recent years. However, the spatiotemporal block tensor method commonly used in this method exhibits large frame spans and oversampling in infrared image sequences. This limitation directly leads to significant interference from clutter in the detection results and poor real-time performance of the entire detection process. Summary of the Invention

[0004] The purpose of the present invention is to propose a method for detecting small targets in infrared images that no longer uses a block tensor method but directly decomposes the image frame. This method mainly improves the real-time and accuracy defects of the existing technology; by combining preliminary decomposition with complete decomposition, the detection results ultimately have higher accuracy and lower false alarm rate.

[0005] The present invention adopts the following technical solutions to achieve the purpose:

[0006] A small target detection method based on a decomposition model is proposed. The method transforms the infrared small target detection task into a background and target decomposition task based on principal component analysis of infrared images. The method in the decomposition task includes the following steps:

[0007] S1. Process the infrared image using the preliminary decomposition model and determine the potential targets in the infrared image by constructing a tensor model;

[0008] S2. Based on the entropy matching method, the motion of the potential target is monitored and the motion change information between the corresponding image frames of the potential target is calculated;

[0009] S3. For an image frame in which the motion change information of a potential target reaches a preset threshold, calculate the weight tensor that needs to be fully decomposed in this image frame, and activate the fully decomposition model at the same time;

[0010] S4. Use the complete decomposition model to process the weight tensor calculated in step S3 to obtain a decomposition result, which is the detection result of the infrared dim small target detection task.

[0011] Furthermore, the tensor model constructed to represent the preliminary decomposition task is:

[0012]

[0013] In the tensor model, is the balance factor, is the original tensor data structure; after solving the tensor model, the target tensor can be determined , which corresponds to the potential target in the infrared image.

[0014] Furthermore, the tensor model is solved by using the alternating direction multiplier method (ADMM). First, the augmented Lagrangian function of the tensor model is determined as follows:

[0015]

[0016] In the functional form, is the Lagrangian operator, is the penalty factor; after the function is alternately minimized, the background tensor is updated and the target tensor , thus realizing the solution process.

[0017] Furthermore, when determining the target tensor After finding the corresponding potential target, the sparse light spot of the potential target's corresponding image is used as the center point to calculate the entropy value in the local range, thereby exploring the motion change information between image frames; the local entropy value with the potential target as the center is defined as follows:

[0018]

[0019]

[0020] Where, The input image block is obtained within a preset specific range with the center point of the potential target as the coordinate point; is the grayscale value, The gray value in the image block is The number of pixels, is the total number of pixels in this image block; The gray value is The ratio of the number of pixels in the total number of pixels in the image block is If Defined as 0.

[0021] Furthermore, for image frames that reach a preset threshold, the inter-frame motion change information is first quantized into the inter-frame standard deviation as the motion feature tensor;

[0022] Then, the target tensors of forward, horizontal and lateral decomposition are obtained by performing t-SVD strategy on the image frame from different sides of the third-order tensor ; Decompose the target tensor into three directions Combined with the motion feature tensor, we get the weight tensor that needs to be fully decomposed .

[0023] Furthermore, the complete decomposition model is a reweighted three-way decomposition model; this model is divided into a forward decomposition model and a horizontal and lateral decomposition model; among them, the forward decomposition model is defined as:

[0024]

[0025] In the formula, the operator Represents the product of two tensors; is the original tensor data structure of the forward solution result, is the forward decomposition part of the background tensor in the image frame; the alternating multiplier method ADMM is used to solve the above formula, thereby updating the sparse tensor of the forward decomposition in the weight tensor .

[0026] The horizontal and lateral decomposition models are defined as:

[0027]

[0028] Where, is the original tensor data structure determined during horizontal and lateral solutions, is the horizontal and lateral decomposition part of the background tensor in the image frame; the alternating multiplier method ADMM is also used to solve the above formula and update the sparse tensor of the horizontal and lateral decomposition ; Combined with sparse tensors , sparse tensor and motion feature matrix , according to the three-dimensional index grouping calculation method, the weight matrix of the corresponding index is obtained ; Weight matrix for all image frames Then reconstruct the weight tensor in index order , as the decomposition result, the background and target decomposition task of principal component analysis of infrared images is completed.

[0029] In summary, due to the adoption of this technical solution, the beneficial effects of the present invention are as follows:

[0030] The method of the present invention reduces the number of channels in the third-order tensor data structure, reducing the overhead of the background and target separation process. The decomposition phase is further refined into two parts: preliminary decomposition and complete decomposition. The preliminary decomposition phase aims to identify potential targets with sparse features in the infrared sequence. The resulting sparse image of the potential target undergoes motion analysis to determine whether any qualified motion has occurred. If motion occurs, the current tensor structure is completely decomposed, using a reweighted tensor approach. This optimization process results in detection results with higher accuracy and lower false alarm rates.

[0031] Several major features of the method of the present invention are:

[0032] 1. The algorithm used takes advantage of the good inter-frame similarity of infrared sequences and directly uses image frames to construct the tensor data structure. This approach can effectively reduce the data structure redundancy caused by oversampling of the spatiotemporal block tensor model and reduce the dimension of the tensor data structure, thereby effectively optimizing the efficiency of the decomposition algorithm and the real-time performance of the processing process.

[0033] 2. The algorithm used divides the entire decomposition process into two stages, further reducing hardware consumption. The initial decomposition and complete decomposition processes use different decomposition models to constrain the sparse tensor with varying tolerances. The initial decomposition prepares for subsequent activation and decomposition, while the complete decomposition uses a less tolerant model to constrain the target, increasing detection accuracy.

[0034] 3. In the motion monitoring stage, the method of the present invention proposes an entropy matching method, which takes the potential target with higher brightness in the sparse tensor as the center point to represent its motion characteristics in the sequence; this method can effectively reduce the consumption of the motion estimation process and is more in line with the characteristics of sparse tensors.

[0035] 4. In the complete decomposition phase, the method proposed in this paper proposes a directional singular value decomposition strategy, t-SVD, as a basic three-way decomposition model to represent the sparse features of the infrared sequence. This is combined with the infrared sequence motion features obtained in the motion monitoring phase to form a weight tensor, which constrains the model's representation of the sparse tensor. This approach can further achieve more accurate decomposition and detection results. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 Schematic diagram of the principle flow of the method of the present invention;

[0037] Figure 2 Schematic diagram of the operation flow of the method of the present invention;

[0038] Figure 3 Schematic diagram of the first-side matrix results of the sparse tensor decomposed in different directions;

[0039] Figure 4 RGB schematic diagram of the three-way decomposition result;

[0040] Figure 5 is a schematic diagram of an example of a weight matrix image;

[0041] Figure 6 This is a schematic diagram of the workflow of the reweighted three-way decomposition model;

[0042] Figure 7 Schematic diagram of a portion of frames of infrared image sequence data 1;

[0043] Figure 8 Schematic diagram of detection results of infrared image sequence data 1;

[0044] Figure 9 Schematic diagram of a portion of frames of infrared image sequence data 2;

[0045] Figure 10 Schematic diagram of the detection results of infrared image sequence data 2. DETAILED DESCRIPTION

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0047] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but rather merely represents selected embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort shall fall within the scope of protection of the present invention.

[0048] Example

[0049] like Figure 1 As shown in FIG, a small target detection method based on a decomposition model is proposed. This method transforms the infrared small target detection task into a background and target decomposition task based on principal component analysis of infrared images. The method in the decomposition task includes the following steps:

[0050] S1. Process the infrared image using the preliminary decomposition model and determine the potential targets in the infrared image by constructing a tensor model;

[0051] S2. Based on the entropy matching method, the motion of the potential target is monitored and the motion change information between the corresponding image frames of the potential target is calculated;

[0052] S3. For an image frame in which the motion change information of a potential target reaches a preset threshold, calculate the weight tensor that needs to be fully decomposed in this image frame, and activate the fully decomposition model at the same time;

[0053] S4. Use the complete decomposition model to process the weight tensor calculated in step S3 to obtain a decomposition result, which is the detection result of the infrared dim small target detection task.

[0054] This embodiment will provide a detailed introduction to each part of the above method steps. Figure 2 The important judgment and processing nodes during the method operation are given and can be referenced simultaneously.

[0055] 1. Preliminary decomposition model.

[0056] In order to transform the infrared dim target detection task into the background and target decomposition task of infrared image principal component analysis, it is necessary to represent the image background and target separately. By constructing a tensor model, the initial decomposition task is expressed as:

[0057]

[0058] in, is the background tensor The nuclear norm is calculated as follows:

[0059]

[0060] In the formula, the tensor is the background tensor Singular value tensor after singular value decomposition; is the Tubal rank of the tensor; this formula represents the singular tensor The sum of the non-zero elements of the first matrix.

[0061] in, is the target tensor of Norm, which is defined as follows:

[0062]

[0063] This formula represents the target tensor The decomposition model based on this definition has a high tolerance for the sparse part in the tensor structure, and will effectively retain the target and some sparse clutter, so that the relevant algorithm can extract the motion information of the sequence based on the retained sparse clutter.

[0064] Furthermore, in the tensor model, is the balance factor, is the original tensor data structure; after solving the tensor model, the target tensor can be determined , which corresponds to the potential target in the infrared image.

[0065] The tensor model is solved using the alternating direction multiplier method (ADMM). First, the augmented Lagrangian function of the tensor model is determined as follows:

[0066]

[0067] In the functional form, is the Lagrangian operator, is the penalty factor;

[0068] The expression for the balancing factor is as follows:

[0069]

[0070] 、 、 That is, it is a single dimension in the three-dimensional direction of the image; the size of the balance factor has a greater impact on the retention of the sparse part of the model. After the function undergoes alternating minimization operations, the background tensor is updated. and the target tensor , thus realizing the solution process.

[0071] In this embodiment, when updating the background tensor and the target tensor In the process of :

[0072]

[0073] This formula has a closed-form solution similar to the soft threshold operator, which is called the tensor singular value contraction operator. Apply fast Fourier transform in three dimensions to get the corresponding tensor , and for the tensor Each face matrix is ​​subjected to the singular value decomposition strategy for the matrix, and the obtained singular value matrix is ​​subjected to the soft threshold operation and reconstructed into a tensor again. , and apply the inverse fast Fourier transform to get the updated background tensor ,Right now .

[0074] Involve the updated background tensor in the target tensor The updates include:

[0075]

[0076] For the Lagrangian operator ,have:

[0077]

[0078] For the penalty factor ,have:

[0079]

[0080] Where, is a calculation parameter value greater than 1, is the maximum penalty factor value of the tensor model.

[0081] 2. Motion monitoring based on entropy matching method.

[0082] The target tensor obtained in the first part The target and some sparse clutter are retained, collectively referred to as potential targets. These potential targets often have higher grayscale values, allowing for clearer positioning and sampling. Based on this, these sparse "light spots" are used as center points to calculate the entropy within their local range, exploring inter-frame motion changes and effectively expressing inter-frame motion information. Furthermore, matching performed in this way only requires calculations for potential targets that meet certain conditions, significantly reducing the system's matching overhead. Furthermore, by reflecting on frames with significant motion changes during subsequent monitoring, the fully decomposed model can be activated.

[0083] In this section, the local entropy value centered on the potential target is defined as follows:

[0084]

[0085]

[0086] Where, The input image block is obtained within a preset specific range with the center point of the potential target as the coordinate point; is the grayscale value, The gray value in the image block is The number of pixels, is the total number of pixels in this image block; The gray value is The ratio of the number of pixels in the total number of pixels in the image block is If Defined as 0.

[0087] Entropy reflects the richness of information contained in an information element. When applied to an image block, it reflects its structural complexity. Generally speaking, the higher the entropy value, the more structural information (such as edges and clutter) it contains; the lower the entropy value, the more continuous and smooth the background. Potential targets left behind in the previous section are often part of this structural information. Image blocks centered around these targets show a clearer entropy value.

[0088] After matching the entropy values ​​of these image blocks, the motion data of the potential target at the corresponding center can be effectively obtained. This data is not only used to test whether there is enough motion between frames, but also the basis for selecting frames to construct the tensor data structure. Therefore, for multiple target tensors Corresponding to multiple potential targets, let the total number of potential targets be , determine the preset threshold value for the entropy value change , used to determine the degree of motion change information.

[0089] 3. Weight tensor for complete decomposition.

[0090] The high-tolerance initial decomposition model is insufficient to remove sparse clutter and hard edges from the sparse tensor. Considering the requirements of infrared search and tracking systems for small target detection rate and false detection rate, the subsequent part uses a complete decomposition model based on the directional characteristics of the t-SVD strategy to obtain more accurate detection results.

[0091] For reaching the preset threshold For image frames, the inter-frame motion change information is first quantified into the inter-frame standard deviation as the motion feature tensor; the process of quantizing the inter-frame motion change information into the inter-frame standard deviation adopts the following formula:

[0092]

[0093] Where, and are the height and width of the image frame respectively, is the motion feature matrix.

[0094] Then, by applying the t-SVD strategy to the image frame from different sides of the third-order tensor, we can obtain sparse tensors with different performances. Figure 3 According to the singular values ​​of different surfaces, they are named forward, horizontal and lateral decomposition. The RGB image composed of the decomposition results of these three directions can more intuitively show their differences, such as Figure 4 As shown. For the target and other sparse clutter in the sparse tensor, the RGB image shows different color characteristics. Based on this characteristic, this embodiment combines the results of the three-directional decomposition with the motion feature tensor to obtain the weight tensor , defined as:

[0095]

[0096] Where, is the number of channels of the image frame; 、 and are the sparse tensors of the forward, horizontal, and lateral decomposition results, respectively.

[0097] An example of the weight matrix image obtained according to this part of the process is shown in the figure below: Figure 5 As shown. Using this weight tensor As the forward decomposition model in Part 4 for sparse tensors The pixel-by-pixel weights of can further constrain the update process to obtain more accurate results.

[0098] 4. Reweighted three-way decomposition model.

[0099] In this embodiment, the complete decomposition model is used to complete the calculation of the complete decomposition stage and obtain the final detection results. The model is a reweighted three-way decomposition model and is divided into a forward decomposition model and a horizontal and lateral decomposition model. The forward decomposition model is defined as:

[0100]

[0101] In the formula, the operator Represents the product of two tensors; is the original tensor data structure of the forward solution result, is the forward decomposition part of the background tensor in the image frame; this model is also solved by the alternating direction multiplier method ADMM, and its augmented Lagrangian function is as follows:

[0102]

[0103] In this embodiment, the solution is performed in the following order: ,have:

[0104]

[0105] For sparse tensors ,have:

[0106]

[0107] For the motion feature matrix involving three parts ,have:

[0108]

[0109] Motion feature matrix Sparse tensor based on forward decomposition The result is calculated, but the weight tensor is further calculated The horizontal and lateral decomposition results need to be used, so the horizontal and lateral decomposition model must be introduced first. The horizontal and lateral decomposition model is defined as:

[0110]

[0111] Where, is the original tensor data structure determined during horizontal and lateral solutions, is the horizontal and lateral decomposition part of the background tensor in the image frame; the alternating direction multiplier method ADMM is also used to solve the above formula, and its augmented Lagrangian function form is:

[0112]

[0113] For the background tensor and To update:

[0114]

[0115] Further sparse tensor and To update:

[0116]

[0117] After obtaining the updated results in all directions in the above round of complete decomposition process, the weight tensor can be updated :

[0118]

[0119] In this embodiment, the weight tensor is updated in each iteration based on the sparse tensor obtained by decomposing in three directions. It has the characteristics of re-weighting and is more in line with the constraints of the model on the target. Moreover, it is only used in the constraints of the forward decomposition results. This approach can make the difference between the forward decomposition result and the decomposition results in the other two directions greater, which is in line with the definition of the weight tensor calculated based on the standard deviation. The calculation process of the first face matrix of the weight tensor in each round of the complete decomposition model is as follows: Figure 6 shown.

[0120] The corresponding sparse tensors are obtained by performing horizontal and lateral decomposition respectively. and Afterwards, the sparse tensor of the forward decomposition result is combined , and the calculated motion feature matrix , according to the three-dimensional index grouping calculation method, the weight matrix of the corresponding index is obtained ; Weight matrix for all image frames Then reconstruct the weight tensor in index order , as the decomposition result, the background and target decomposition task of principal component analysis of infrared images is completed.

[0121] The above four parts are the entire contents of the method of this embodiment. The following introduces the results of infrared image dim target detection in related experiments.

[0122] Figure 7 Given an infrared image sequence data 1, Figure 8 This is the detection result for Data 1. The infrared image sequence Data 1 has a rapidly changing background and contains a lot of clutter and hard edges. This makes detection quite challenging. However, the results show that the model retains the target well overall. Hard edges and sparse clutter are minimally retained in the resulting image, and are present at a very low pixel level. This indicates that the model's false alarm rate is also relatively good.

[0123] Figure 9 Given infrared image sequence data 2 with a more complex background, the detection results are Figure 10 ; Under the same detection test, the test results show that the method of this embodiment can still stably detect the target.

Claims

1. A small target detection method based on a decomposition model, characterized by: The infrared dim target detection task is transformed into a background and target decomposition task based on the principal component analysis of infrared images. The decomposition method includes the following steps: S1. Process the infrared image using the preliminary decomposition model and determine the potential targets in the infrared image by constructing a tensor model; S2. Based on the entropy matching method, the motion of the potential target is monitored and the motion change information between the corresponding image frames of the potential target is calculated; S3. For an image frame in which the motion change information of a potential target reaches a preset threshold, calculate the weight tensor that needs to be fully decomposed in this image frame, and activate the fully decomposition model at the same time; S4. Use the complete decomposition model to process the weight tensor calculated in step S3 to obtain a decomposition result, which is the detection result of the infrared dim small target detection task; For multiple target tensors Corresponding to multiple potential targets, let the total number of potential targets be , determine the preset threshold value for the entropy value change , used to determine the degree of motion change information; In step S3, for the For image frames, the inter-frame motion change information is first quantified into the inter-frame standard deviation as the motion feature tensor; Then, the target tensors of forward, horizontal and lateral decomposition are obtained by performing t-SVD strategy on the image frame from different sides of the third-order tensor ; Decompose the target tensor into three directions Combined with the motion feature tensor, we get the weight tensor that needs to be fully decomposed ; The process of quantifying the inter-frame motion change information into the inter-frame standard deviation is as follows: Where, and are the height and width of the image frame respectively, is the motion feature matrix; The weight tensor obtained by combining related tensors is defined as follows: Where, is the number of channels of the image frame; 、 and are the sparse tensors of the forward, horizontal and lateral decomposition results respectively.

2. The small target detection method based on a decomposition model according to claim 1, characterized in that: In step S1, the tensor model constructed to represent the initial decomposition task is: In the tensor model, is the background tensor The nuclear norm is calculated as follows: In the formula, the tensor is the background tensor Singular value tensor after singular value decomposition; is the Tubal rank of the tensor; this formula represents the singular tensor The sum of the non-zero elements of the first matrix; In the tensor model, is the target tensor of Norm, which is defined as follows: This formula represents the target tensor The sum of the absolute values ​​of all non-zero elements in; Furthermore, in the tensor model, is the balance factor, is the original tensor data structure; After solving the tensor model, the target tensor can be determined , which corresponds to the potential target in the infrared image.

3. The small target detection method based on a decomposition model according to claim 2, characterized in that: The tensor model is solved using the alternating direction multiplier method (ADMM). First, the augmented Lagrangian function of the tensor model is determined as follows: In the functional form, is the Lagrangian operator, is the penalty factor; The expression for the balancing factor is as follows: 、 、 That is, the single dimension in the three-dimensional direction of the image; after the function is alternately minimized, the background tensor is updated and the target tensor , thus realizing the solution process.

4. The small target detection method based on a decomposition model according to claim 3, characterized in that: Update the background tensor and the target tensor In the process of ,have: The closed-form solution of this equation is the tensor singular value contraction operator, which is used to shrink the background tensor Apply fast Fourier transform in three dimensions to get the corresponding tensor , and for the tensor Each face matrix is ​​subjected to the singular value decomposition strategy for the matrix, and the obtained singular value matrix is ​​subjected to the soft threshold operation and reconstructed into a tensor again. , and apply the inverse fast Fourier transform to get the updated background tensor ,Right now ; Involve the updated background tensor in the target tensor The updates include: For the Lagrangian operator ,have: For the penalty factor ,have: Where, is a calculation parameter value greater than 1, is the maximum penalty factor value of the tensor model.

5. The method for detecting small and dim targets based on a decomposition model according to claim 3, wherein: In step S2, after determining the target tensor After finding the corresponding potential target, the sparse light spot of the potential target's corresponding image is used as the center point to calculate the entropy value in the local range, thereby exploring the motion change information between image frames; the local entropy value with the potential target as the center is defined as follows: Where, The input image block is obtained within a preset specific range with the center point of the potential target as the coordinate point; is the grayscale value, The gray value in the image block is The number of pixels, is the total number of pixels in this image block; The gray value is The ratio of the number of pixels in the total number of pixels in the image block is If Defined as 0.

6. The method for detecting small and dim targets based on a decomposition model according to claim 1, wherein: In step S4, the complete decomposition model is a reweighted three-way decomposition model; this model is divided into a forward decomposition model and a horizontal and lateral decomposition model; wherein the forward decomposition model is defined as: In the formula, the operator Represents the product of two tensors; is the original tensor data structure of the forward solution result, is the forward decomposition part of the background tensor in the image frame; the alternating multiplier method ADMM is used to solve the above formula, thereby updating the sparse tensor of the forward decomposition in the weight tensor .

7. The method for detecting small and dim targets based on a decomposition model according to claim 6, wherein: The horizontal and lateral decomposition models are defined as: Where, is the original tensor data structure determined during horizontal and lateral solutions, is the horizontal and lateral decomposition part of the background tensor in the image frame; the alternating multiplier method ADMM is also used to solve the above formula and update the sparse tensor of the horizontal and lateral decomposition ; Combined with sparse tensors , sparse tensor and motion feature matrix , according to the three-dimensional index grouping calculation method, the weight matrix of the corresponding index is obtained ; Weight matrix for all image frames Then reconstruct the weight tensor in index order , as the decomposition result, the background and target decomposition task of principal component analysis of infrared images is completed.

Images