An aircraft infrared dim small target detection method and system based on trajectory constraint tensor decomposition

Abstract

The application discloses an aircraft infrared dim small target detection method and system based on trajectory constraint tensor decomposition, and belongs to the field of infrared image processing.The method comprises the following steps: acquiring infrared sequence images of an aircraft to be detected, and constructing an infrared sequence tensor of the aircraft; constructing an infrared sequence tensor decomposition model according to the characteristics of the infrared sequence images of the aircraft; solving the aircraft motion trajectory obtained by the infrared sequence tensor model, and constructing an aircraft target motion trajectory perception matrix; constructing a trajectory constraint tensor of an aircraft target component of the tensor decomposition model based on the target motion state perception ability of the motion trajectory perception matrix; solving the improved infrared sequence tensor decomposition model, acquiring an aircraft target detection confidence map, and constructing an adaptive detection threshold to realize the target detection of the aircraft.The application can improve the accuracy of aircraft target detection and tracking based on infrared sequence images.

Description

Technical Field

[0001] This invention relates to the field of image processing, and in particular to a method and system for detecting weak infrared targets of aircraft based on trajectory constraint tensor decomposition. Background Technology

[0002] With the continuous advancement of drone technology and the growth of market demand, small drone technology has developed rapidly, and related regulatory and safety issues have become increasingly prominent. The widespread use of small drones has raised a series of concerns related to privacy, security, and airspace management. Against this backdrop, target detection and tracking of small drones has become a crucial issue. The development of this technology can be used to monitor drone activity, provide early warnings of potential threats, ensure flight safety, and also help prevent violations, protect privacy, and maintain public safety. Effective detection and tracking systems can help regulatory agencies and private companies implement better drone management, reduce potential risks, and ensure public safety and privacy.

[0003] Patent CN114973143A, a robust detection method for low-altitude aircraft based on motion features, proposes a method for detecting small low-altitude drones. However, the Gaussian background modeling method is insufficient to robustly represent the complex background of infrared image sequences. Furthermore, the size range of the drones to be detected is limited to 20-30 mm, making it unsuitable for real-world detection requirements. Pre-trained low-altitude aircraft target detection models struggle to acquire sufficient infrared image sequence data tagged with small drones. Therefore, this patent application aims to provide an innovative infrared target detection method for small drones to address the regulatory and safety needs of this emerging field. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method for detecting small infrared targets of aircraft based on trajectory constraint tensor decomposition, which can quickly detect small targets of aircraft and achieve rapid early warning in video sequences captured by infrared remote sensing cameras.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] This invention proposes a method for detecting weak infrared targets of aircraft based on trajectory-constrained tensor decomposition, comprising the following steps:

[0007] (1) Input the infrared sequence image of the aircraft to be detected and construct the infrared sequence tensor of the aircraft;

[0008] (2) Based on the sparse characteristics of the infrared weak target of the aircraft, as well as the low-rank characteristics of the background and the temporal dimension structure characteristics of the noise, an infrared sequence tensor decomposition model is constructed. The infrared sequence tensor is decomposed using the infrared sequence tensor decomposition model to obtain the background component, target component and noise component of the infrared sequence tensor.

[0009] (3) The target components obtained from the infrared sequence tensor decomposition model at past moments are used as the aircraft's motion trajectory. Based on the aircraft's motion trajectory, the aircraft target motion trajectory perception matrix is ​​constructed.

[0010] (4) Construct the trajectory constraint tensor of the target component in the infrared sequence tensor decomposition model based on the target motion trajectory sensing matrix of the aircraft;

[0011] (5) Based on the trajectory constraint tensor of the target component, an improved infrared sequence tensor decomposition model is constructed, the improved infrared sequence tensor decomposition model is solved, the aircraft target detection confidence map is obtained, and an adaptive detection threshold is constructed to realize the detection of the aircraft target.

[0012] Furthermore, the infrared sequence tensor of the aircraft is:

[0013]

[0014] Where t represents the current time, and ΔT represents the time-series sensing window scale of the custom infrared sequence tensor. This represents the infrared image at time t-k+1 in the infrared image sequence. This represents the k-th forward slice of the infrared sequence tensor, Δt. The infrared sequence tensor at time t

[0015] Furthermore, the infrared sequence tensor decomposition model is as follows:

[0016]

[0017] in ε t Let the background component, target component, and noise component of the infrared sequence tensor and the infrared sequence tensor be represented respectively. This means finding the function that minimizes the given information. ε t , The tensor low-rank constraint represents the background component, ||·||1 represents the L1 sparse norm, and ||·|| 1,1,2 L represents 1,1,2 Norms, λ1,λ2>0, represent the tensor decomposition weights.

[0018] Furthermore, the low-rank constraint of the tensor of the background component is:

[0019]

[0020] Where α i The pattern expansion weight, α i >0, Indicates background components Perform tensor pattern expansion i, i = 1, 2, 3, ||·|| * Represents the matrix nuclear norm.

[0021] Furthermore, the target motion trajectory perception matrix of the aircraft is as follows:

[0022]

[0023] Where S t This represents the sensing matrix of the aircraft target motion trajectory at time t, where γ>0 represents the weight attenuation coefficient. This represents the first forward slice of the target component at time k obtained by solving the above infrared sequence tensor decomposition model.

[0024] Furthermore, the trajectory constraint tensor of the target component in the infrared sequence tensor decomposition model is:

[0025]

[0026]

[0027] in This represents the k-th forward slice of the motion trajectory constraint tensor at time t. S represents the identity matrix. t-k+1 This represents the sensing matrix of the aircraft target motion trajectory at time t-j+1 constructed in step 3), where ΔT represents the temporal sensing window scale of the custom infrared sequence tensor; ΔT elements The trajectory constraint tensor at time t

[0028] Furthermore, the improved tensor decomposition model is as follows:

[0029]

[0030] Where ⊙ represents the Hadamarda.

[0031] Furthermore, the constructed aircraft target detection confidence map is as follows:

[0032]

[0033] Among them, O t This represents the target detection confidence map of the aircraft at time t. This represents the first forward slice of the target component at time t obtained by solving the improved infrared sequence tensor decomposition model.

[0034] Furthermore, the adaptive detection threshold is:

[0035] τ t =mean(O t )+β×std(O t )

[0036] Where τ t denoted by t, the adaptive detection threshold is defined by mean(.), std(.) represents the mean function, std(.) represents the standard deviation function, and β represents the empirical parameter. If the aircraft target detection confidence value is higher than the detection threshold, then the point is determined to be the location of the detected target.

[0037] This invention also proposes an infrared weak target detection system for aircraft to implement the above-mentioned aircraft infrared weak target detection method based on trajectory constraint tensor decomposition, which includes:

[0038] The infrared sequence image processing module is used to acquire infrared sequence images of the aircraft to be detected and construct the infrared sequence tensor of the aircraft.

[0039] The infrared sequence tensor decomposition module is used to construct and solve the infrared sequence tensor decomposition model to obtain the background component, target component and noise component of the infrared sequence tensor.

[0040] The target motion trajectory constraint module is used to receive the target components of the past time obtained by the infrared sequence tensor decomposition model, construct the target motion trajectory perception matrix of the aircraft, and construct the trajectory constraint tensor of the target components in the infrared sequence tensor decomposition model based on the target motion trajectory perception matrix of the aircraft.

[0041] The detection module receives the trajectory constraint tensor of the target components obtained by the target motion trajectory constraint module, constructs and solves the improved infrared sequence tensor decomposition model, obtains the aircraft target detection confidence map, and constructs an adaptive detection threshold to realize the detection of the aircraft target.

[0042] Compared with the prior art, the present invention has the following advantages:

[0043] (1) The present invention constructs an infrared sequence tensor based on infrared sequence images, which makes full use of the spatial and temporal dimensions of the data. The constructed tensor decomposition model effectively represents the background, target and noise characteristics, and improves the detection and recognition accuracy.

[0044] (2) The trajectory constraint tensor of the present invention can effectively perceive the historical motion trajectory of the aircraft target. Identifying the historical motion trajectory of the aircraft's weak infrared target is the key to distinguishing the target from structural noise. The trajectory constraint tensor fully characterizes the non-consistency of the aircraft target's temporal dimension, making the target component of the tensor decomposition model pay more attention to the motion attributes of the aircraft target and suppressing the expression of structural noise, thereby improving the accuracy of tensor decomposition and further enhancing the detection and recognition accuracy.

[0045] (3) The infrared sequence tensor decomposition model of the present invention accurately separates the infrared weak target from the background and noise, effectively making up for the problem that the existing detection model is difficult to effectively detect weak targets in complex environments, and improving the robustness of infrared weak target detection and tracking of aircraft. Attached Figure Description

[0046] Figure 1 This is a flowchart of the present invention;

[0047] Figure 2 This is an infrared sequence image of an aircraft target captured by an infrared camera in one embodiment of the present invention;

[0048] Figure 3 This is a diagram showing the infrared detection results of a small target in an aircraft according to an embodiment of the present invention. Detailed Implementation

[0049] To more clearly illustrate the objectives, technical solutions, and advantages of this invention, we will provide a detailed description in conjunction with specific implementation examples. The following sections include descriptions of specific implementation examples to simplify the understanding of the invention. It should be emphasized that the invention is not limited to the described implementation examples, and various modifications can be made without departing from its basic principles. These equivalent variations are also included within the scope of the claims of this patent application.

[0050] See Figure 1 The present invention provides a method for detecting aircraft targets based on trajectory constraint tensor decomposition, comprising:

[0051] Step 1) Obtain the infrared sequence image of the aircraft to be detected and construct the infrared sequence tensor.

[0052] In this embodiment, the acquired infrared image sequence includes an air-to-ground background and an aircraft target, such as... Figure 2 As shown, the image size is 128×128. Considering the spatial correlation between the ground background and the airspace in a single frame of infrared image sequence, and the strong temporal consistency of infrared images at adjacent times in an infrared image sequence, directly constructing the infrared sequence tensor of the aircraft under test based on the infrared image sequence can preserve the spatial and temporal dimensions of the data structure.

[0053]

[0054] in Represents the infrared sequence tensor at time t The k-th forward slice, Let represent a single frame of infrared image at time t-k+1, where m and n represent the length and width of the single frame, and ΔT represents the time-aware window scale of the custom infrared sequence tensor. Specifically, in this embodiment, ΔT is set to 5.

[0055] Step 2) Based on the sparse characteristics of the aircraft's weak infrared targets, as well as the low-rank characteristics of the image background and the temporal dimensional structure characteristics of the noise, an infrared sequence tensor decomposition model is constructed. The infrared sequence tensor is decomposed using the infrared sequence tensor decomposition model to obtain the background component, target component and noise component of the infrared sequence tensor.

[0056] Specifically, based on the spatial dimensional correlation and temporal dimensional consistency of the image background, the Tucker tensor rank is improved, assigning different weight coefficients to the kernel norm of different modes to represent the tensor background components. Based on the spatial dimensional sparsity and temporal dimensional inconsistency of the infrared weak targets of the aircraft, the L1 sparse norm is used to represent the tensor target components. Based on the distribution characteristics of Gaussian noise and salt-and-pepper noise, L... 1,1,2 The norm characterizes the noise component of a tensor. Considering the different component characteristics of the tensor mentioned above, a tensor decomposition model for infrared sequences is constructed:

[0057]

[0058] in ε t Let represent the infrared sequence tensor at time t respectively. Background components, target components, and noise components. This means finding the function that minimizes the given information. ε t rank(·) denotes the low-rank constraint of the tensor, ||·||1 denotes the L1 sparse norm, and ||·|| 1,1,2 L represents 1,1,2 The norm, λ1,λ2>0, represents the tensor decomposition weights. In this specific embodiment, the values ​​of λ1 and λ2 are both 0.5.

[0059] The aforementioned low-rank tensor constraint is specifically as follows:

[0060]

[0061] Where α i The pattern expansion weights are α1>α2, α1>α3, and α i >0, Indicates background components Perform tensor pattern expansion i, i = 1, 2, 3, ||·|| * The matrix norm is represented by α1. Considering that different expansion modes of the infrared sequence tensor can express data features of different dimensions, mode 1 expansion focuses on temporal dimension features, while modes 2 and 3 expansions focus on spatial dimension features. Furthermore, considering that the main differences between the target and background, and structural noise, lie in the inconsistency of the temporal dimension, a larger weight is assigned to the matrix norm of mode 1 expansion. Specifically, in this embodiment, α1, α2, and α3 are set to 0.8, 0.1, and 0.1, respectively.

[0062] Step 3) Use the target components obtained from the infrared sequence tensor decomposition model at past moments as the aircraft's motion trajectory, and construct the aircraft target motion trajectory perception matrix based on the aircraft's motion trajectory.

[0063] Specifically, the moving target of an aircraft manifests as directional positional changes in the infrared sequence tensor. By comprehensively considering the historical information of the target components in the tensor, the tensor decomposition model can be effectively guided to focus on the target's motion status, thus improving the accuracy of the tensor decomposition. The constructed aircraft target motion trajectory perception matrix is ​​as follows:

[0064]

[0065] Where S t This represents the sensing matrix of the aircraft target motion trajectory at time t, where γ>0 represents the weight attenuation coefficient. This represents the target component obtained at time k from the above infrared sequence tensor decomposition model. The first forward slice. In this specific embodiment, γ is set to 0.8.

[0066] Step 4) Based on the target motion state perception capability of the aircraft target motion trajectory perception matrix, construct the trajectory constraint tensor of the aircraft target component in the infrared sequence tensor decomposition model.

[0067] Specifically, the infrared sequence tensor decomposition model of the above formula (2) has the problem of difficulty in suppressing structured noise. Based on the target perception capability of the motion trajectory perception matrix, introducing trajectory constraint tensors to the aircraft target components can improve the robustness of tensor decomposition. The trajectory constraint tensors of the aircraft target components are constructed as follows:

[0068]

[0069] in This represents the k-th forward slice of the motion trajectory constraint tensor at time t. S represents the identity matrix. t-k+1 This represents the sensing matrix of the aircraft target motion trajectory at time t-k+1 constructed in step 3), with ΔT forward slices. The trajectory constraint tensor at time t

[0070] Step 5) Construct and solve the improved infrared sequence tensor decomposition model, obtain the aircraft target detection confidence map, and construct an adaptive detection threshold to realize the detection of aircraft targets.

[0071] The improved infrared sequence tensor decomposition model is as follows:

[0072]

[0073] Where ⊙ represents the Hadamarda.

[0074] Specifically, the improved infrared sequence tensor decomposition model is an NP-hard problem, and an analytical solution cannot be obtained directly. This invention uses the alternating Lagrange operator method to solve for the target components.

[0075] Furthermore, based on the effective enhancement of the flying target in the target component, the first forward slice in the target component is extracted as the aircraft target detection confidence map at time t:

[0076]

[0077] Among them, O t This represents the target detection confidence map of the aircraft at time t. This represents the first forward slice of the target component at time t obtained by solving the improved infrared sequence tensor decomposition model described above.

[0078] Based on the confidence map of aircraft target detection, an adaptive detection threshold τ is constructed. t As shown in formula (8), the point in the aircraft target detection confidence map where the aircraft target detection confidence value is higher than the detection threshold is determined as the location of the detected target.

[0079] τ t =mean(O t )+β×std(O t (8)

[0080] O t >τ t (9)

[0081] Where mean(.) represents the mean function, std(.) represents the standard deviation function, and β represents the empirical parameter. Specifically, in the implementation example... Figure 3 The infrared detection results for small targets on aircraft show that background and noise components are suppressed, while target components are significantly enhanced.

[0082] The accompanying drawings vividly illustrate the purpose, technical solutions, and advantages of this invention, further highlighting its clarity. It should be particularly emphasized that these specific examples are intended only to explain the principles of the invention and not to limit its scope. Any equivalent substitutions or improvements that conform to the methodological ideas and principles provided by this invention should be included within the scope of patent protection of this invention.

Claims

1. A method for detecting weak infrared targets of aircraft based on trajectory-constrained tensor decomposition, characterized in that, Includes the following steps: (1) Input the infrared sequence image of the aircraft to be detected and construct the infrared sequence tensor of the aircraft; (2) Construct an infrared sequence tensor decomposition model, and use the infrared sequence tensor decomposition model to decompose the infrared sequence tensor to obtain the background component, target component and noise component of the infrared sequence tensor; (3) The target components obtained from the infrared sequence tensor decomposition model at past moments are used as the aircraft's motion trajectory. Based on the aircraft's motion trajectory, the aircraft target motion trajectory perception matrix is ​​constructed. The target motion trajectory sensing matrix of the aircraft is: ； in This represents the sensing matrix of the aircraft target motion trajectory at time t. This represents the weight decay coefficient. This indicates that the solution obtained from the above infrared sequence tensor decomposition model is in The first forward slice of the target component at time step; (4) Construct the trajectory constraint tensor of the target component in the infrared sequence tensor decomposition model based on the target motion trajectory perception matrix of the aircraft; The trajectory constraint tensor of the target component in the infrared sequence tensor decomposition model is: ； in express The first moment of the motion trajectory constraint tensor A forward slice, Represents the identity matrix. This indicates the construction in step 3). The spacecraft target motion trajectory perception matrix at any given time. The temporal sensing window scale represents a custom infrared sequence tensor; indivual composition Trajectory constraint tensor at time step ; (5) Based on the trajectory constraint tensor of the target component, construct an improved infrared sequence tensor decomposition model, solve the improved infrared sequence tensor decomposition model, obtain the aircraft target detection confidence map, and construct an adaptive detection threshold to realize the detection of the aircraft target.

2. The method according to claim 1, characterized in that, The infrared sequence tensor of the aircraft is: , ； in Indicates the current moment. The temporal sensing window scale represents a custom infrared sequence tensor. Indicates the first in an infrared sequence image Infrared image at any moment, The first tensor representing the infrared sequence tensor A forward slice, indivual The infrared sequence tensor at time t .

3. The method according to claim 1, characterized in that, The infrared sequence tensor decomposition model is as follows: ； in Let the background component, target component, and noise component of the infrared sequence tensor and the infrared sequence tensor be represented respectively. This means finding the function that minimizes the given information. rank( ) represents the low-rank constraint of the background component tensor. express Sparse norm, express Norm, This represents the tensor decomposition weights.

4. The method according to claim 3, characterized in that, The tensor low-rank constraint of the background component is: ； in Indicates the pattern expansion weight. , Indicates background components Perform tensor pattern expansion for i, where i = 1, 2, 3. Represents the matrix nuclear norm.

5. The method according to claim 1, 3, or 4, characterized in that, The improved tensor decomposition model for infrared sequences is as follows: ； in It represents the Hadamah accumulation. Let represent the infrared sequence tensor at time t respectively. Background components, target components, and noise components. This means finding the function that minimizes the given information. , express Sparse norm, express Norm, Indicates background components Perform tensor pattern expansion for i, where i = 1, 2, 3. Denotes the nuclear norm of a matrix. Indicates the pattern expansion weight. This represents the tensor decomposition weights.

6. The method according to claim 1, characterized in that, The aforementioned aircraft target detection confidence map is as follows: ； in, express Confidence map for target detection of aircraft at any time. This represents the solution obtained from the improved infrared sequence tensor decomposition model. The first forward slice of the target component at time step.

7. The method according to claim 6, characterized in that, The adaptive detection threshold is: ； in Let represent the adaptive detection threshold at time t, mean(.) denote the mean function, and std(.) denote the standard deviation function. This represents an empirical parameter; if the aircraft target detection confidence value is higher than the adaptive detection threshold, then the point is determined to be the location of the detected target.

8. A system for detecting weak infrared targets of an aircraft, used to implement the method for detecting weak infrared targets of an aircraft based on trajectory constraint tensor decomposition as described in claim 1, characterized in that, It includes: The infrared sequence image processing module is used to acquire infrared sequence images of the aircraft to be detected and construct the infrared sequence tensor of the aircraft. The infrared sequence tensor decomposition module is used to construct and solve the infrared sequence tensor decomposition model to obtain the background component, target component and noise component of the infrared sequence tensor. The target motion trajectory constraint module is used to receive the target components of the past time obtained by the infrared sequence tensor decomposition model, construct the target motion trajectory perception matrix of the aircraft, and construct the trajectory constraint tensor of the target components in the infrared sequence tensor decomposition model based on the target motion trajectory perception matrix of the aircraft. The detection module receives the trajectory constraint tensor of the target components obtained by the target motion trajectory constraint module, constructs and solves the improved infrared sequence tensor decomposition model, obtains the aircraft target detection confidence map, and constructs an adaptive detection threshold to realize the detection of the aircraft target.

Images