A method for infrared small target detection based on multi-angle prior analysis

Abstract

The present invention discloses a method for detecting small infrared targets based on multi-angle prior analysis, and the specific steps are as follows: Step 1, preprocessing the original image to obtain a grayscale image, and constructing a slice tensor model of the original image; Step 2, filtering the grayscale image through its corresponding directional derivative model, and constructing a corresponding slice tensor; Step 3, adding local structural weights to the grayscale image; Step 4, separating the target from the background; Step 5, image reconstruction and target detection. The present invention proposes a multi-directional prior analysis, and enhances the target by a re-weighted method, so as to achieve accurate detection of targets in infrared images in harsh environments lacking texture information and spatial structure information, thereby reducing the false alarm rate. The method focuses on the balance between suppressing the background and enhancing the target during small target detection, that is, while obtaining and suppressing the background edge, the target edge information is retained.

Description

Technical Field

[0001] The invention relates to the technical field of infrared small target detection, in particular to an infrared small target detection method based on multi-angle priori analysis. Background Art

[0002] Infrared images have a long imaging distance, excellent anti-interference performance, and the advantage of being able to obtain good images at any time of the day. As an important means of photoelectric detection, infrared imaging has been widely used in space surveillance, remote sensing, medical imaging, target detection and tracking. However, the performance of these applications depends largely on robust target detection. However, since the infrared imaging system is far away from the target, the target is often presented as a point in the infrared image, and its distance information cannot be displayed. On the one hand, the target in the infrared image is very small, and the specific shape and texture cannot be collected, and obvious spatial structure information cannot be used. On the other hand, due to the complex background and large imaging noise, the signal-to-noise ratio of infrared images is usually very low, making target detection extremely difficult. In addition, target detection in infrared images is often interfered by various interferences such as the edge of thick clouds, sea clutter, and artificial heat sources on the ground, which will cause a high false alarm rate and weaken the detection performance. Therefore, studying infrared small target detection under complex backgrounds is a very valuable and challenging task.

[0003] At present, the methods based on infrared small target detection can be divided into two types: single-frame based detection and sequence-based detection. Among them, the sequence-based method obtains the target's movement trajectory by using the consistent information of adjacent frames and using spatiotemporal information to suppress the interference of noise and clutter. The sequence-based detection method improves the detection performance of infrared small targets. Since the sequence-based detection method is highly dependent on spatiotemporal information and requires a slow relative motion between the imaging device and the target, in practical applications, it is difficult for the imaging device to maintain the uniformity and consistency of the background, and the movement of the target is random, which will lead to inconsistent information between adjacent frames, resulting in a high false alarm rate. Unlike sequence-based detection, the single-frame-based detection method is often used in high-speed moving target detection, real-time detection and other platforms due to its advantages such as low hardware requirements and fast processing time. Moreover, the results of single-frame detection can also be applied to sequence-based detection, providing higher performance for sequence-based detection. Since the single-frame detection scheme lacks spatiotemporal information, it is particularly important to efficiently utilize the characteristics of the target and background in the single-frame image. Therefore, in recent years, more and more scholars have begun to pay attention to the single-frame detection method. Whether the prior information can be used efficiently determines the performance of single-frame target detection.

[0004] How to efficiently detect small targets in complex background clutter has become a research difficulty in the field of infrared small target detection. Some scholars have introduced the principal component analysis method into the field of small target detection, and transformed the infrared small target detection task from separating targets to classifying targets and backgrounds. For example, in the prior art, a Chinese patent application with application publication number CN109447073A discloses a method for infrared weak small target detection based on tensor robust principal component analysis, including constructing a structural weight tensor, constructing an objective function using tensor robust principal component analysis, inputting a third-order tensor and a structural weight tensor into the objective function, solving the objective function using the alternating direction multiplier method to obtain background tensors and target tensors, reconstructing background images and target images, performing adaptive threshold segmentation on the target image to determine the position of the target, and outputting the target detection result. This method achieves the effect of improving target detection and background suppression capabilities and enhancing target shape retention capabilities. In addition, Gao et al. proposed the infrared patch image model (IPI, Infrared Patch-Image), which converts the original infrared image into the original matrix, and then uses partially overlapping image blocks and sliding windows to increase the correlation of the background, and finally recovers the sparse matrix and low-rank matrix from the original matrix. The core of this method is to assume that the background comes from a single low-rank subspace, regard the target as an outlier, and transform the infrared small target detection problem into a low-rank sparse optimization problem of the target sparse prior, thereby achieving the separation of background and target. Compared with the traditional filter-based detection method, the IPI model shows superior background suppression performance. However, the IPI model also has disadvantages such as excessive target shrinkage, clutter residual, and long time consumption. Recently, Dai et al. improved the construction method of patch images, introduced the concept of tensor, and proposed the infrared patch tensor (IPT, Infrared Patch-Tensor) model based on the IPI model. This model assumes that all matrix models are low-rank, thereby improving the detection ability and greatly reducing the calculation time. In the IPT model for single-frame detection proposed by Dai et al., they used the two highest eigenvalues ​​in the structure tensor to describe the local geometric structure, but when using λ1-λ2 to determine the edge, they could not distinguish whether the edge belongs to the target or the background. This ambiguity directly leads to the distortion of the target shape in the small target detection results. In addition, the IPT model only considers the background-related priors, and does not consider the target-related priors. Summary of the invention

[0005] In view of the above problems, the present invention proposes an infrared small target detection method based on multi-angle prior analysis on the basis of the IPT model. Under low signal-to-noise ratio, the target can be detected more accurately and quickly, and the false alarm rate caused by background clutter can be reduced.

[0006] In order to achieve the above object, the present invention is implemented by the following technical solutions:

[0007] The present invention is an infrared small target detection method based on multi-angle prior analysis, comprising the following steps:

[0008] Step 1: Convert the RGB image to a grayscale image and construct a slice tensor of the original image;

[0009] Step 2: Filter the grayscale image through its corresponding directional derivative model and construct the corresponding slice tensor;

[0010] Step 3: Add local structure weight to the grayscale image to further distinguish targets and sparse edges;

[0011] Step 4: Input the fragment tensor obtained in step 1 and step 2 and the local structure weight obtained in step 3 into the tensor part and norm model for solution, and iterate through the alternating direction multiplication method to obtain the fragment tensor for separating the target and the background;

[0012] Step 5: Restore the obtained target and background fragment tensors into images and extract small target information.

[0013] A further improvement of the present invention is that in step 1, a picture is cut and superimposed according to the size of the sliding window 40×40 to form a three-dimensional tensor of 40×40×t, where t is the number of sliding windows.

[0014] A further improvement of the present invention is that the pixel surface function g(r, c) in step 2 is expressed as:

[0015]

[0016]

[0017] The first-order directional derivative of g(r,c) at any angle α is calculated as:

[0018]

[0019] Among them, P i is a two-dimensional discrete Chebyshev polynomial, k i is the coefficient of the two-dimensional discrete Chebyshev polynomial, and I(r, c) represents the original pixel value. In the specific experiment of the present invention, α is taken as 0°, 30°, 60°, and -60° respectively. The original image is processed by the directional derivative g′ α The derivative feature graphs in these four directions can be obtained respectively. A further improvement of the present invention is that: the local structure weight W in step 3 sm M sm The inverse of the multi-directional edge-aware mapping M sm It is expressed as:

[0020]

[0021] Where j∈{1, 2, 3, …, n} represents n different directions. For each direction, the derivative feature map obtained has its corresponding structure tensor E:

[0022]

[0023] Represented as the gradient of the image, It is represented as the transpose of the gradient of the image. x and I y Respectively represent the horizontal gradient and vertical gradient of the image. Since the structure tensor E is a symmetric semi-positive two-dimensional matrix, the two eigenvalues ​​are:

[0024]

[0025] The local structure perception mapping in a certain j direction is:

[0026]

[0027] λ j1 ,λ j2 represents the eigenvalue of the corresponding structure tensor in a certain j direction. When one eigenvalue is much larger than another eigenvalue (approaching 0), it means that the image has a significant edge structure, so the difference in feature rank is replaced by the maximum value max(λ between the two eigenvalues) j1 ,λ j2 ).

[0028] A further improvement of the present invention is that in step 5, an adaptive threshold segmentation algorithm is used to extract small targets, and pixels in the image that are greater than a threshold value Th are target pixels, and the threshold value is as follows:

[0029] Th=max(d min , μ+ρσ)

[0030] where μ and σ are the mean and standard deviation of the reconstructed target image, respectively. min and ρ are constants set empirically.

[0031] The beneficial effects of the present invention are:

[0032] (1) The present invention uses a method based on multi-angle prior analysis to enhance the target, solving the problem that the introduction of local priors causes the target to shrink to varying degrees. The traditional local structure weight can easily use the difference between the two features of the structure tensor to determine the edge component, obtain edge information, and highlight small targets by removing edge information. However, the local structure weight will also highlight the edge information of the target, which means that the target is reduced to a certain extent, resulting in lower accuracy. Since the target area changes from bright to dark along the derivative direction, the edge characteristics are obvious, and the local transformation of background clutter is relatively gentle, so the targets at different angles have stronger characteristics than the clutter in the background. Therefore, the present invention introduces a local structure descriptor related to the target prior in the local derivatives in multiple directions, which is expressed as half of the harmonic mean of the two eigenvalues. When one eigenvalue is much larger than the other eigenvalue, it means that there is a significant edge structure. Therefore, replacing the difference between the two eigenvalues ​​with the maximum value between the two eigenvalues ​​can improve the presence of the target edge while removing the background clutter, overcoming the problem of missing target edge information when suppressing the background edge.

[0033] (2) The present invention uses multi-angle prior analysis to construct weights and re-weights sparse targets in the overall model, so that the model can not only classify the target and the background and better protect the shape of the original target, thereby improving the detection accuracy, but also accelerate the convergence speed and alleviate the imbalance that smaller values ​​are subject to greater penalties. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is the overall framework diagram of the model of the present invention.

[0035] Figure 2 This is the original image of the infrared image dataset.

[0036] Figure 3 This is the target result diagram after detection by the present invention.

[0037] Figure 4 It is a 3D potential energy diagram of the detection result diagram of the present invention.

[0038] Figure 5 The experimental results of the present invention are compared and displayed. DETAILED DESCRIPTION

[0039] The following will be combined with the accompanying drawings in the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the implementation described is only a part of the implementation of the present invention, not all of the implementations. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0040] The present invention is an infrared small target detection method based on multi-angle prior analysis, and the specific steps are as follows: Step 1: Preprocess the original image: Convert the RGB image to a grayscale image. In order to explore the association between different image slices, cut and superimpose a picture according to the size of a sliding window of 40×40 to form a three-dimensional tensor of 40×40×t, that is, obtain the original slice tensor Wherein t is the number of sliding windows; in the specific experiment of the present invention, the value of t is 40.

[0041] Step 2: Filter the grayscale image through its corresponding directional derivative model to achieve the purpose of enhancing the target and suppressing background clutter; and construct the corresponding fragment tensor;

[0042] The solution of the directional derivative of the small neighborhood in the present invention is specifically: use a binary cubic function to fit the 5×5 neighborhood, take the symmetric set R={-2,-1,0,1,2}, and the discrete orthogonal multinomial set is The symmetric set of columns C = {-2, -1, 0, 1, 2}, and its corresponding discrete orthogonal polynomial set is Then, ignoring the terms of polynomials higher than 3rd order, 10 two-dimensional discrete Chebyshev polynomials are constructed.

[0043] The pixel surface function g(r, c) of this neighborhood is:

[0044]

[0045] Among them, i ranges from 1 to 10, I(r, c) represents the original pixel value, and the coefficient k can be simplified according to the orthogonality of the polynomial. i Calculation of coefficient k i It is expressed as:

[0046]

[0047] Among them, because I(r, c) is independent of other terms, the coefficient k i It can be regarded as a fixed filter w of the original image i , expressed as:

[0048]

[0049] The first-order directional derivative of g(r, c) at any angle α can be calculated using the following formula:

[0050]

[0051] After the image is filtered by the first-order directional derivative, a 40×40 window is slid from the upper left to the lower right. The number of window slides is t, and the corresponding slice tensor of the derivative feature map is obtained.

[0052] Step 3: Local prior extraction, adding local structure weight to the grayscale image to further distinguish targets and sparse edges;

[0053] The derivative feature map obtained in each direction in step 2 has its corresponding structure tensor E:

[0054]

[0055] Represented as the gradient of the image, It is represented as the transpose of the gradient of the image. x and I y Respectively represent the horizontal gradient and vertical gradient of the image. Since the structure tensor E is a symmetric semi-positive two-dimensional matrix, two non-negative eigenvalues ​​λ1 and λ2 can be calculated. The calculation formula is:

[0056]

[0057] Set four directions and map LE through local perception j (x, y) to obtain the local prior intensity map in four directions:

[0058]

[0059] In the specific experiment of the present invention, the four directions are set to 0°, 30°, 60°, and -60° respectively. j∈{1, 2, 3, 4} represents the four different directions, λ j1 , j2 The two non-negative eigenvalues ​​calculated for the structure tensor in the four directions.

[0060] Based on the fusion strategy, a multi-directional edge-aware mapping can be introduced, namely:

[0061]

[0062] And M sm The reciprocal of constitutes the local structure weight W sm The local structure weights are added by sliding a 40×40 window from the upper left to the lower right. Convert to local structure weight tensor

[0063] Step 4: Input the slice tensor obtained in steps 1 and 2 and the local structure weight obtained in step 3 into the tensor part and norm model for solution. The augmented Lagrangian function of the specific model can be written as:

[0064]

[0065] in is the target shard tensor; is the background slice tensor, is the local structure weight tensor, is the Lagrange multiplier; β is a tradeoff parameter; μ is a penalty factor greater than 0; <·> represents the inner product of two third-order tensors; is represented as the rank of the background tensor; ||·||1 is the l1 norm, which is the sum of the absolute values ​​of all items; ||·|| F is the Frobenius norm; ⊙ represents the Hadamard product. In the specific experiment of the present invention, μ=0.003, β=0.015.

[0066] Then, using the alternating direction multiplication method, The minimization of is decomposed into several sub-problems, and Can be updated through iterative methods.

[0067] Step 4.1 Target shard tensor Convert its subproblem into the following function:

[0068]

[0069] By taking the derivative of the augmented Lagrangian function and setting it equal to 0, we can deduce The update function for the k+1th time is as follows:

[0070]

[0071] The above function can be solved by the following equation:

[0072]

[0073] Where S τ (·) is a soft threshold operator.

[0074]

[0075] Where sign(x) is the sign function.

[0076] Step 4.2: For the background slice tensor Can be converted to the following function:

[0077]

[0078] By taking the derivative of the augmented Lagrangian function and setting it equal to 0, we can deduce The update function for the k+1th time is as follows:

[0079]

[0080] Step 4.3: Guan Ding Updates;

[0081] Taking the derivative of the augmented Lagrangian and setting it equal to 0, we can deduce The update function for the k+1th time is as follows:

[0082]

[0083] Step 4.4 Update of μ

[0084] μ k+1 =cμ k

[0085] Wherein c is a constant greater than 1. In the specific experiment of the present invention, c=1.1.

[0086] By iterating the above steps, we can obtain the fragment tensor of target and background separation, that is, we can obtain the low-rank background fragment tensor and the sparse target shard tensor

[0087] Step 5: perform image reconstruction and target detection;

[0088] Step 5.1 sequentially extracts t front slices of the slice tensor and reconstructs the two-dimensional image using the inverse operation of the tensor constructed in step 1. Background f B and target f T Can be respectively through the background slice tensor and the target shard tensor Reconstruction acquisition.

[0089] Step 5.2 Considering that the pixels of the real target in the reconstructed target image have a high grayscale, a simple adaptive threshold segmentation algorithm can be used to extract small targets. The threshold is as follows:

[0090] Th=max(d min ,μ+ρσ) (7)

[0091] Where μ and σ are the mean and standard deviation of the reconstructed target image respectively. min and ρ are constants set empirically. When a pixel in the reconstructed target image is greater than Th, it is the target pixel.

[0092] The present invention detects the infrared small target image, and the results are as follows Figure 3 and Figure 4 As shown, Figure 3 The result diagram of the detection is shown in Figure 2. Figure 4 As shown in the resulting 3D image, it can be clearly seen that the present invention can well suppress background clutter and enhance targets to achieve the function of accurately detecting small targets.

[0093] In order to verify the effectiveness of the present invention, a comparative experiment was conducted, specifically: IPI (infrared patch image model), VAR_DIFF, RLCM, NARM, SRWS and NOLC test methods were used to test on the same data set, and compared with the experimental results of the present invention. The test comparison results are shown in Tables 1, 2 and Figure 5 As shown, Figure 5 Figures a, b, c, and d of each algorithm correspond to the data sets Figure 2 a, b, c, d in the above.

[0094] Table 1 shows the SCRG values ​​of various comparison methods

[0095] IPI NARM RLCM VAR_DIFF SRWS NOLC OURS a 2.98 15.42 15.70 23.59 25.94 10.26 23.65 b 2.65 32.44 21.00 76.90 25.96 11.69 147.47 c 6.42 3.15 3.03 7.10 3.55 3.08 8.09 d 4.14 22.89 16.62 27.09 21.23 10.63 24.16

[0096] Table 2 shows the BSF value of each method.

[0097] IPI NARM RLCM VAR_DIFF SRWS NOLC OURS a 2.46 12.73 6.68 22.96 21.42 8.47 22.55 b 2.53 26.85 5.91 73.36 20.89 9.00 82.21 c 6.42 2.88 2.14 7.04 3.51 3.06 8.04 d 4.14 22.90 4.15 26.18 21.23 10.62 23.29

[0098] Combining Table 1, Table 2 and Figure 5 From the comparison results, it can be seen that the effect of the present invention is better than that of the compared algorithms in terms of both suppressing the background and enhancing the target.

[0099] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for detecting small infrared targets based on multi-angle prior analysis, characterized in that: The steps include: Step 1: Convert the original RGB image to a grayscale image and construct a slice tensor of the original image; Step 2: Filter the grayscale image through its corresponding directional derivative model and construct the corresponding slice tensor; Step 3: Add local structure weights containing target priors to the grayscale image to further distinguish targets from sparse edges; Step 4: Input the fragment tensor obtained in step 1 and step 2 and the local structure weight obtained in step 3 into the tensor part and norm model for solution, and iterate through the alternating direction multiplication method to obtain the fragment tensor for separating the target and the background; Step 5: Restore the obtained target and background fragment tensors into images and extract small target information; The local structure weight W in step 3 sm M sm The inverse of the multi-directional edge-aware mapping M sm It is expressed as: Where j∈{1, 2, 3, ..., n} represents n different directions, and the image obtained in each direction has its corresponding structure tensor E: Represented as the gradient of the image, Represented as the transpose of the gradient of the image, where I x and I y Respectively represent the horizontal gradient and vertical gradient of the image, and two non-negative eigenvalues ​​λ1 and λ2 are calculated through the structure tensor E; since the structure tensor E is a symmetric semi-positive two-dimensional matrix, the two eigenvalues ​​are: Local Aware Mapping (LE) in Multi-Directional Edge Aware Mapping j (x, y) is as follows: In the formula, λ j1 ,λ j2 are the two non-negative eigenvalues ​​of the structure tensor E in a specific j direction.

2. The infrared small target detection method based on multi-angle prior analysis according to claim 1 is characterized in that: In step 1, an image is cut and superimposed according to the size of the sliding window 40×40 to form a three-dimensional tensor of 40×40×t, where t is the number of sliding windows.

3. The infrared small target detection method based on multi-angle prior analysis according to claim 2 is characterized in that: In step 2, the grayscale image is filtered by the corresponding first-order directional derivative and then the corresponding fragment tensor is constructed. The solution process of the first-order directional derivative is as follows: The surface function g(r, c) at a specific pixel (r, c) is expressed as: Where r and c represent the corresponding coordinate values ​​of the pixel point in the small area, and g(r, c) is the first-order directional derivative g′ at any angle α. α Calculated as: Among them, P i is a two-dimensional discrete Chebyshev polynomial, k i are the two-dimensional discrete Chebyshev polynomial coefficients, I(r, c) represents the original pixel value, Expressed as partial derivative.

4. The infrared small target detection method based on multi-angle prior analysis according to claim 1 is characterized in that: In step 5, an adaptive threshold segmentation algorithm is used to extract small targets. The pixels in the image that are greater than the threshold Th are target pixels. The threshold is as follows: Th=max(d min ,m+rσ) where μ and σ are the mean and standard deviation of the reconstructed target image, respectively. min and p are constants set empirically.

5. A computer device, comprising at least one computing device, characterized in that: The computing device includes a memory, a processor, and a program configuration stored in the memory for executing an infrared small target detection method based on multi-angle prior analysis according to any one of claims 1 to 4.

Images