Infrared single small target detection method and system based on local difference and global feature

Abstract

The application discloses an infrared single small target detection method and system based on local difference and global features, applied to the field of infrared imaging technology, and comprises the following steps: constructing a local difference sliding window, and designing an efficient local difference operator; the whole image is traversed pixel by pixel through the sliding window, and a local difference feature matrix is obtained according to the local difference operator; the threshold value of the local difference matrix is calculated, and the position of a candidate target is obtained through the threshold value; the global uniqueness feature of each candidate target point is calculated, and a global uniqueness feature matrix is obtained; the local difference matrix and the global uniqueness matrix are normalized respectively, and then are multiplied, so that a fusion matrix is obtained; and the infrared small target is obtained by carrying out binaryzation segmentation on the fusion matrix through an adaptive threshold value. The application comprehensively utilizes the local difference and the uniqueness of the single target, effectively suppresses background clutter, improves the detection performance of the infrared single small target, and has the advantages of simple principle, low complexity and easy realization.

Description

Technical Field

[0001] This invention relates to the fields of image processing and target detection technology, and specifically to an infrared single small target detection method and system based on local differences and global features. Background Technology

[0002] Infrared small target detection is widely used in fields such as homeland surveillance, maritime monitoring, air defense early warning, and missile defense. It is a crucial component of infrared search and track systems, and its detection results directly impact the robustness of these systems. However, it faces numerous challenges, including small target size, irregular shape, weak signal strength, and lack of texture detail. Therefore, achieving high robustness and low false alarm rate detection is a pressing issue for researchers.

[0003] To date, researchers have proposed numerous single-frame detection algorithms, which can be broadly categorized into four types. The first type is based on background estimation filtering, which performs poorly in complex scenes. The second type is based on low-rank sparsity methods, which have achieved better progress in background suppression, but their ability to suppress corners and sparse edges still needs improvement, and their high complexity leads to poor real-time performance. The third type is based on deep learning methods, which have strong target representation capabilities but heavily rely on training databases. Currently, although some public datasets exist, they cannot cover all complex scenes, making deep learning algorithms lack generalization ability in complex environments. The fourth type is based on local information methods, which utilize local features to enhance targets and suppress clutter. Local feature-based methods are favored by researchers due to their significant target enhancement effects, good real-time performance, and ease of hardware implementation. Researchers have improved algorithm performance by developing various local contrast measurements and different local window structures. To fully extract local features from small targets, they have attempted different technical approaches, such as using random walks and region growing algorithms to segment small targets. However, when clutter has local features similar to or stronger than those of small targets, such algorithms may mistake it for a target, leading to false alarms. These algorithms rely solely on the local features of small targets, facing the dilemma of not being able to further improve the detection rate and reduce the false alarm rate. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide an infrared single small target detection method and system based on local differences and global features. It takes into account the non-local characteristics of infrared images and the uniqueness of single targets, and quantifies the feature into a global uniqueness parameter that can be used to enhance the target while suppressing high-brightness clutter. By weightedly fusing this parameter with local features, the infrared single small target detection performance can be further improved and the false alarm rate can be reduced.

[0005] The technical solution adopted in this invention is: an infrared single small target detection method based on local differences and global features, comprising the following steps:

[0006] Step 1: Obtain a single frame of the original infrared image to be detected, in which only one small target exists;

[0007] Step 2: Construct a multi-scale local difference sliding window, and design an efficient local difference operator based on the sliding window structure. The local difference operator calculates the difference between the gray value of the center pixel and the gray value of the largest pixel in the background area at each scale to measure the local features.

[0008] Step 3: Fill the original infrared image by traversing the entire image pixel by pixel using the sliding window from Step 2, and calculate the local difference feature value of each pixel according to the local difference operator to obtain a local difference feature matrix with the same size as the original image.

[0009] Step 4: Calculate the threshold of the local difference matrix, and use this threshold to obtain the positions of m candidate target points;

[0010] Step 5: Calculate the global uniqueness features of each candidate target point and obtain the global uniqueness feature matrix;

[0011] Step 6: Normalize the local difference matrix and the global uniqueness matrix respectively, and then perform dot product to obtain the fusion matrix;

[0012] Step 7: Calculate the adaptive threshold of the fusion matrix, and use this threshold to perform binarization segmentation of the fusion matrix to obtain infrared small targets.

[0013] Furthermore, the multi-scale local difference sliding window W has multiple scales, where each single-scale local difference sliding window W S The size is S×S, where S refers to the outer side length of the window at different scales. The single-scale local difference sliding window includes the center pixel, the background area with a certain pixel width around it, and the transition area between the center pixel and the background area.

[0014] Furthermore, for each pixel in the infrared image, a local difference sliding window centered on that pixel is constructed, and then the single-scale local difference ld of that pixel is defined as follows:

[0015]

[0016] Where T(x,y) and B max (x, y) represent the pixel values ​​in the original infrared image and the maximum pixel value of the background region B, respectively;

[0017] The local difference operator designed based on the multi-scale sliding window W is expressed as follows:

[0018]

[0019] Where LD(x,y) represents the local difference feature value of a pixel, and S=7,9,11 refers to the sliding window with side lengths of 7, 9 and 11 pixels respectively.

[0020] Furthermore, the specific implementation method of step 4 is as follows;

[0021] First, the dilation matrix P of the local difference feature matrix LD is calculated using morphological dilation. Then, the region where the dilation matrix and the local difference feature matrix have the same gray value is defined as the local maximum region. A zero-matrix Z with the same size as the original image is constructed, and the values ​​of the corresponding local maximum regions in Z are assigned to 1. The mathematical formula is as follows:

[0022]

[0023] Where LD(x,y) and P(x,y) refer to the element values ​​at coordinates (x,y) on the local difference feature matrix LD and the expansion matrix P, respectively. Then, Z is multiplied by the local difference feature matrix LD to obtain the local maximum feature matrix ZL. Then, the elements in ZL with an intensity higher than the adaptive threshold Th are selected. ZL The m elements are used as candidate targets, and their coordinates are stored. An adaptive threshold Th is applied. ZL The calculation formula is as follows:

[0024] Th ZL =max(α×V1,V k )

[0025] Where V1 and V k α represents the maximum value and the k-th largest pixel value of the local maximum feature matrix ZL, where α and k are empirical constants.

[0026] Furthermore, the formula for calculating the expansion matrix P is as follows:

[0027] P=LD⊕J

[0028] Where ⊕ refers to the grayscale expansion operation, and J represents the structured element that performs the expansion operation on the local difference feature matrix.

[0029] Furthermore, the specific implementation method of step 5 is as follows;

[0030] First, a grayscale histogram is introduced to distinguish between the target and clutter. For the k-th candidate target point, a local pixel block is first constructed on the original infrared image, centered on the coordinates of the k-th candidate target. Then, the grayscale histogram features of the local pixel block are extracted. A weighting function is introduced to give more weight to pixels near the target center, ultimately obtaining the grayscale distance-weighted histogram of the local pixel block of the k-th candidate target point. bin refers to the total number of histogram intervals;

[0031] Then, the gray-level histogram calculation operation is repeated for all candidate target points to obtain an m×m histogram matrix. To measure the similarity between different histograms, the Bach coefficient is used for similarity calculation. The weighted histogram p of the k-th candidate target point is then calculated. k The weighted histogram p of the j-th candidate target point j The Bartholomew's coefficient ρ(p) between k ,p j The mathematical formula for ) is as follows:

[0032]

[0033] After calculating the similarity coefficient between the weighted histogram of each candidate target point and the weighted histograms of other candidate target points, an m×m similarity matrix Mat is obtained. B In this process, the similarity value between each local pixel block and itself is reassigned to 0; a zero-matrix GU with the same size as the original image is constructed to store the global uniqueness feature values, where the element values ​​of the candidate target coordinates are calculated using the following formula:

[0034]

[0035] Where max(θ) k (Refers to the similarity matrix Mat) B The maximum value of the elements in the k-th column, x k and y k The coordinates of the k-th candidate target point are given. After calculating the global uniqueness eigenvalues ​​of all candidate target points, the global uniqueness eigenvalue matrix GU is obtained.

[0036] Furthermore, The calculation formula is as follows:

[0037]

[0038] Where C refers to the normalization coefficient, N refers to the number of pixels in the local pixel block, d refers to the distance kernel function, h refers to the bandwidth of the kernel function, z0 refers to the coordinates of the center pixel of the local pixel block, and z i The coordinates of the i-th pixel within a local pixel block are given by δ, the Dirac function is given by bin, and the total number of histogram intervals is given by b(z). i ) indicates that z i The grayscale value of a pixel is mapped to the index value of the histogram interval;

[0039] Furthermore, the specific implementation method of step 7 is as follows;

[0040] In the fusion matrix R, the true target is extracted from the candidate target using an adaptive threshold, which is defined as follows:

[0041] T=λ×R max +(1-λ)×R mean

[0042] Where R max and R mean λ is the maximum and mean of the fusion matrix R, and λ is a given parameter. After thresholding, pixel values ​​greater than or equal to the threshold in the fusion matrix are set to 1, and pixel values ​​less than the threshold are set to 0, resulting in a binary image. Pixels with a value of 1 will be used as the target output.

[0043] This invention also discloses an infrared single small target detection system based on local differences and global features, comprising the following modules:

[0044] The raw infrared image acquisition module is used to acquire a single frame of raw infrared image to be detected, in which only a small target exists;

[0045] The local difference operator construction module is used to construct a multi-scale local difference sliding window. Based on the sliding window structure, an efficient local difference operator is designed. The local difference operator calculates the difference between the gray value of the center pixel and the gray value of the largest pixel in the background area at each scale to measure local features.

[0046] The local difference feature matrix acquisition module is used to fill the original infrared image. It traverses the entire image pixel by pixel through the sliding window, calculates the local difference feature value of each pixel according to the local difference operator, and obtains a local difference feature matrix with the same size as the original image.

[0047] The threshold calculation module for the local difference matrix is ​​used to calculate the threshold of the local difference matrix and obtain the positions of m candidate target points through this threshold.

[0048] The global uniqueness feature matrix acquisition module is used to calculate the global uniqueness features of each candidate target point and obtain the global uniqueness feature matrix.

[0049] The fusion matrix acquisition module is used to normalize the local difference matrix and the global uniqueness matrix respectively, and then perform a dot product to obtain the fusion matrix.

[0050] The infrared small target acquisition module is used to calculate the adaptive threshold of the fusion matrix, and then use this threshold to perform binarization segmentation of the fusion matrix to obtain the infrared small target.

[0051] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0052] This invention considers the nonlocal characteristics of infrared image background and the uniqueness of single targets, quantifying this feature into parameters that can be used to enhance targets and suppress high-brightness clutter. Furthermore, it proposes an efficient local difference feature extraction operator. By utilizing the local difference features and global uniqueness features of small infrared targets, it further improves single-target detection performance, overcoming the problem of existing detection methods struggling to suppress a large amount of high-brightness clutter when detecting small targets in complex scenes. Simultaneously, the computation process of this invention is simpler and easier to implement in hardware. Attached Figure Description

[0053] Figure 1 A flowchart provided for an embodiment of the present invention;

[0054] Figure 2 This is a schematic diagram of the local difference window structure provided in an embodiment of the present invention;

[0055] Figure 3 This is a schematic diagram of similarity matrix calculation provided in an embodiment of the present invention;

[0056] Figure 4 The detection results of the method provided in the embodiments of the present invention and the existing algorithms. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0058] like Figure 1 As shown, the target detection method process is represented by a flowchart.

[0059] The specific steps of the infrared single small target detection method based on local differences and global features described in this invention are as follows:

[0060] Step 1: Obtain the original single-frame infrared image O to be detected. Its size is M×N. The gray value of the pixel in image O is represented as O(x,y), where x and y represent the pixel coordinates, x=1,2,…,M, y=1,2,…,N. There is only one small target in the image.

[0061] Step 2, construct a multi-scale local difference sliding window W, the structure of which is as follows: Figure 2 As shown, W has three scales, where each single-scale local difference sliding window W SThe size of the window is S×S, where S refers to the outer side length of the window at different scales. In the image, S is 7, 9, and 11 pixels at the three scales, respectively. It consists of the center pixel (represented in red) and the surrounding 2-pixel wide background region B (represented in blue). The transition region between the target pixel and the background region (represented in white) is 1, 2, and 3 pixels wide, respectively. Then, an efficient local difference operator is designed based on this window structure. The local difference operator calculates the difference between the gray value of the center pixel and the gray value of the largest pixel in the background region at each scale to measure local features. The maximum value of the three differences is defined as the local difference feature value of the current pixel. To effectively enhance the target and suppress most of the background, a local difference operator is designed to describe local features. For each pixel in the infrared image, a local difference sliding window centered on that pixel is constructed, and the single-scale local difference ld of that pixel is defined as follows:

[0062]

[0063] Where T(x,y) and B max (x, y) represent the pixel value in the original infrared image and the maximum pixel value of the neighboring background B, respectively. Since typical detection tasks target bright targets, values ​​less than 0 are set to zero. Small target scales range from 2×2 to 9×9, with the target center pixel value being greater than the target edge pixel value. For example... Figure 2 The three-scale windows shown can effectively detect targets of unknown size. The local difference operator designed based on the multi-scale sliding window W is expressed as follows:

[0064]

[0065] Where S = 7, 9, 11 refers to... Figure 2 The sliding windows shown have side lengths of 7, 9, and 11 pixels, respectively.

[0066] Step 3: Fill the original infrared image by traversing the entire image pixel by pixel using the sliding window from Step 2, and calculate the local difference feature value of each pixel according to the local difference operator to obtain the local difference feature matrix LD with the same size as the original image.

[0067] Step 4: Calculate the threshold of the local difference matrix, and obtain the positions of m candidate target points using this threshold; calculate the expansion matrix P of the local difference feature matrix LD through morphological expansion operation, which can be expressed by the following formula:

[0068] P=LD⊕J

[0069] Where ⊕ refers to the grayscale dilation operation, and J represents the structured element that undergoes the dilation operation on the local difference feature matrix. Then, the region where the grayscale values ​​of the dilation matrix and the local difference feature matrix are consistent is defined as the local maximum region. A zero-matrix Z with the same size as the original image is constructed, and the values ​​of the corresponding local maximum regions in Z are assigned the value of 1. The mathematical formula is as follows:

[0070]

[0071] Where LD(x,y) and P(x,y) refer to the element values ​​at coordinates (x,y) in the local difference feature matrix LD and the expansion matrix P, respectively. Then, Z is multiplied by the local difference feature matrix LD to obtain the local maximum feature matrix ZL. Finally, elements in ZL with intensity higher than the adaptive threshold Th are selected. ZL The m elements are used as candidate targets, and their coordinates are stored. An adaptive threshold Th is applied. ZL The calculation formula is as follows:

[0072] Th ZL =max(α×V1,V k )

[0073] Where V1 and V k α represents the maximum value of the local maximum feature matrix ZL and the k-th largest pixel value, where α and k are empirical constants. In an optional embodiment, the optimal value of α is 0.1, and the optimal value of k is 40.

[0074] Step 5: Calculate the global uniqueness features of each candidate target point and obtain the global uniqueness feature matrix;

[0075] This embodiment takes into account that targets and clutter have different gray-level distribution structures in human visual perception. Therefore, a gray-level histogram is introduced to distinguish targets and clutter in order to measure the difference. Taking the k-th candidate target point as an example, on the original infrared image, an 11×11 local pixel block is first constructed with the coordinates of the k-th candidate target as the center, and then the gray-level histogram features of the local pixel block are extracted. Pixels at the target edge are relatively far from the target center, which can easily mix with background information and interfere with the histogram of the target area. Therefore, a weighting function is introduced to give more weight to pixels near the target center, and finally the gray-level distance-weighted histogram of the local pixel block of the k-th candidate target point is obtained. in The mathematical formula is as follows:

[0076]

[0077] Where C refers to the normalization coefficient, N refers to the number of pixels in the local pixel block, d refers to the distance kernel function, h refers to the bandwidth of the kernel function, z0 refers to the coordinates of the center pixel of the local pixel block, and z iThe coordinates of the i-th pixel within a local pixel block are given by δ, the Dirac function is given by bin, and the total number of histogram intervals is given by b(z). i ) indicates that z i The grayscale value of a pixel is mapped to the index value of the histogram interval. In an optional embodiment, the optimal value of h is 72, and the optimal value of bin is 32.

[0078] Then, this operation is repeated for all candidate target points to obtain an m×m histogram matrix. To measure the similarity between different histograms, the similarity coefficient is calculated using the Bach coefficient, and the weighted histogram p of the k-th candidate target point is used. k The weighted histogram p of the j-th candidate target point j The Bartholomew's coefficient ρ(p) between k ,p j The mathematical formula for ) is as follows:

[0079]

[0080] Where bin refers to the histogram interval number.

[0081] like Figure 3 As shown, after calculating the similarity coefficient between the weighted histogram of each candidate target point and the weighted histograms of other candidate target points, an m×m similarity matrix Mat is obtained. B In this process, the similarity value between each local pixel block and itself is reassigned to 0. A zero-based matrix GU with the same size as the original image is constructed to store the global uniqueness feature values. The formula for calculating the element value GU(x,y) of the candidate target coordinates is as follows:

[0082] Where max(θ) k (Refers to the similarity matrix Mat) B The maximum value of the elements in the k-th column, x k and y k This refers to the coordinates of the k-th candidate target point. After calculating the global uniqueness eigenvalues ​​of all candidate target points, the global uniqueness feature matrix GU can be obtained.

[0083] Step 6: Normalize the local difference matrix and the global uniqueness matrix respectively, and then perform dot product to obtain the fusion matrix R;

[0084] Step 7: Calculate the adaptive threshold T of the fusion matrix R, and use this threshold to perform binarization segmentation of the fusion matrix to obtain the infrared small target.

[0085] In the fusion matrix R, background clutter is suppressed, and pixels with high intensity values ​​are most likely to be the target. Therefore, a true target is extracted from the candidate targets using an adaptive threshold. The adaptive threshold T is defined as follows:

[0086] T=λ×R max +(1-λ)×R mean

[0087] Where R max and R mean λ represents the maximum and mean values ​​of the fusion matrix R, and λ is a given parameter. In one optional embodiment, the optimal range for λ is between 0.5 and 0.7. After thresholding, pixel values ​​greater than or equal to the threshold in the fusion matrix are set to 1, and pixel values ​​less than the threshold are set to 0, resulting in a binary image. Pixels set to 1 will be used as the target output.

[0088] To demonstrate the effectiveness and detection capability of our proposed method, infrared images of small targets in four complex scenes were selected for validation, and the results were compared with other commonly used target detection methods. The other comparison algorithms included: LCM, MPCM, AADCDD, TTLLCM, PLLCM, PSTNN, LogTFNN, and ACM-U-Net. Figure 4 The detection results of various algorithms are shown, where red rectangles indicate detected targets and blue ellipses indicate residual clutter. Figure 4 The comparative algorithms in the previous paper failed to effectively suppress the bright clutter in the four scenes, while the method of this invention detected a single target without any residual background clutter. This demonstrates that the method of this invention has better detection performance and stronger background clutter suppression capability for small infrared targets against complex backgrounds.

[0089] This invention also discloses an infrared single small target detection system based on local differences and global features, comprising the following modules:

[0090] The raw infrared image acquisition module is used to acquire a single frame of raw infrared image to be detected, in which only a small target exists;

[0091] The local difference operator construction module is used to construct a multi-scale local difference sliding window. Based on the sliding window structure, an efficient local difference operator is designed. The local difference operator calculates the difference between the gray value of the center pixel and the gray value of the largest pixel in the background area at each scale to measure local features.

[0092] The local difference feature matrix acquisition module is used to fill the original infrared image. It traverses the entire image pixel by pixel through the sliding window, calculates the local difference feature value of each pixel according to the local difference operator, and obtains a local difference feature matrix with the same size as the original image.

[0093] The threshold calculation module for the local difference matrix is ​​used to calculate the threshold of the local difference matrix and obtain the positions of m candidate target points through this threshold.

[0094] The global uniqueness feature matrix acquisition module is used to calculate the global uniqueness features of each candidate target point and obtain the global uniqueness feature matrix.

[0095] The fusion matrix acquisition module is used to normalize the local difference matrix and the global uniqueness matrix respectively, and then perform a dot product to obtain the fusion matrix.

[0096] The infrared small target acquisition module is used to calculate the adaptive threshold of the fusion matrix, and then use this threshold to perform binarization segmentation of the fusion matrix to obtain the infrared small target.

[0097] The specific implementation methods of each module are the same as those of each step, and will not be described in this invention.

[0098] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An infrared single small target detection method based on local differences and global features, characterized in that, Includes the following steps: Step 1: Obtain a single frame of the original infrared image to be detected, in which only one small target exists; Step 2: Construct a multi-scale local difference sliding window, and design an efficient local difference operator based on the sliding window structure. The local difference operator calculates the difference between the gray value of the center pixel and the gray value of the largest pixel in the background area at each scale to measure the local features. Step 3: Fill the original infrared image by traversing the entire image pixel by pixel using the sliding window from Step 2, and calculate the local difference feature value of each pixel according to the local difference operator to obtain a local difference feature matrix with the same size as the original image. Step 4: Calculate the threshold of the local difference matrix, and use this threshold to obtain the positions of m candidate target points; Step 5: Calculate the global uniqueness features of each candidate target point and obtain the global uniqueness feature matrix; The specific implementation method of step 5 is as follows: First, a grayscale histogram is introduced to distinguish between the target and clutter. For the k-th candidate target point, a local pixel block is first constructed on the original infrared image, centered on the coordinates of the k-th candidate target. Then, the grayscale histogram features of the local pixel block are extracted. The weighting function assigns more weight to pixels near the target center, ultimately resulting in a gray-level distance-weighted histogram of the local pixel block of the k-th candidate target point. , This refers to the total number of histogram intervals; Then, the gray-level histogram calculation is repeated for all candidate target points to obtain an m×m histogram matrix. To measure the similarity between different histograms, the Bach coefficient is used for similarity calculation, and the weighted histogram of the k-th candidate target point is calculated. The weighted histogram of the j-th candidate target point The Bartholomew's coefficient between The mathematical formula is as follows: After calculating the similarity coefficient between the weighted histogram of each candidate target point and the weighted histogram of other candidate target points, we obtain... similarity matrix Mat B In this process, the similarity value between each local pixel block and itself is reassigned to 0; a zero-matrix GU with the same size as the original image is constructed to store the global uniqueness feature values, where the element values ​​of the candidate target coordinates are calculated using the following formula: in Similarity matrix Mat B The maximum value of the elements in the k-th column. and The coordinates of the k-th candidate target point are given. After calculating the global uniqueness eigenvalues ​​of all candidate target points, the global uniqueness eigenvalue matrix GU is obtained. Step 6: Normalize the local difference matrix and the global uniqueness matrix respectively, and then perform dot product to obtain the fusion matrix; Step 7: Calculate the adaptive threshold of the fusion matrix, and use this threshold to perform binarization segmentation of the fusion matrix to obtain infrared small targets.

2. The infrared single small target detection method based on local differences and global features as described in claim 1, characterized in that: In step 2, the multi-scale local difference sliding window W has multiple scales, wherein each single-scale local difference sliding window W S The size is S×S, where S refers to the outer side length of the window at different scales. The single-scale local difference sliding window includes the center pixel, the background area with a certain pixel width around it, and the transition area between the center pixel and the background area.

3. The infrared single small target detection method based on local differences and global features as described in claim 2, characterized in that: For each pixel in the infrared image, a local difference sliding window centered on that pixel is constructed, and then the single-scale local difference of that pixel is defined. as follows: in and These represent the pixel values ​​in the original infrared image and the maximum pixel value of the background region B, respectively. The local difference operator designed based on the multi-scale sliding window W is expressed as follows: in, Represents the local difference feature value of a pixel. These refer to sliding windows with side lengths of 7, 9, and 11 pixels respectively.

4. The infrared single small target detection method based on local differences and global features as described in claim 1, characterized in that: The specific implementation method of step 4 is as follows; First, the dilation matrix P of the local difference feature matrix LD is calculated using morphological dilation. Then, the region where the dilation matrix and the local difference feature matrix have the same gray value is defined as the local maximum region. A zero-matrix Z with the same size as the original image is constructed, and the values ​​of the corresponding local maximum regions in Z are assigned to 1. The mathematical formula is as follows: Among them, and These refer to the local difference feature matrix LD and the expansion matrix, respectively. P The upper coordinate is The element values ​​of Z are then multiplied by the local difference feature matrix LD to obtain the local maximum feature matrix ZL. Then, the elements in ZL with intensity higher than the adaptive threshold Th are selected. ZL The m elements are used as candidate targets, and their coordinates are stored. An adaptive threshold Th is applied. ZL The calculation formula is as follows: in and The maximum value and the th local maximum characteristic matrix ZL represent the maximum value and the th local maximum characteristic matrix ZL. k Large pixel values, and k It is an empirical constant.

5. The infrared single small target detection method based on local differences and global features as described in claim 4, characterized in that: Inflation matrix P The calculation formula is as follows: in This refers to the grayscale expansion operation. J This represents a structured element that has undergone an expansion operation on the local difference feature matrix.

6. The infrared single small target detection method based on local differences and global features as described in claim 1, characterized in that: The calculation formula is as follows: in Refers to the normalization coefficient. The number of pixels within a local pixel block. The distance kernel function. The bandwidth of the kernel function. The coordinates of the center pixel of a local pixel block. The coordinates of the i-th pixel within a local pixel block. The Dirac function, This refers to the total number of histogram intervals. Indicates will The grayscale value of a pixel is mapped to the index value of the histogram interval.

7. The infrared single small target detection method based on local differences and global features as described in claim 1, characterized in that: The specific implementation method of step 7 is as follows; In the fusion matrix R, the true target is extracted from the candidate target using an adaptive threshold, which is defined as follows: in and These are the maximum and mean values ​​of the fusion matrix R. It is a given parameter. After thresholding, the pixel values ​​in the fusion matrix that are greater than or equal to the threshold are set to 1, and the pixel values ​​that are less than the threshold are set to 0, resulting in a binary image. The pixels that are set to 1 will be used as the target output.

8. An infrared single small target detection system based on local differences and global features, characterized in that, Includes the following modules: The raw infrared image acquisition module is used to acquire a single frame of raw infrared image to be detected, in which only a small target exists; The local difference operator construction module is used to construct a multi-scale local difference sliding window. Based on the sliding window structure, an efficient local difference operator is designed. The local difference operator calculates the difference between the gray value of the center pixel and the gray value of the largest pixel in the background area at each scale to measure local features. The local difference feature matrix acquisition module is used to fill the original infrared image. It traverses the entire image pixel by pixel through the sliding window, calculates the local difference feature value of each pixel according to the local difference operator, and obtains a local difference feature matrix with the same size as the original image. The threshold calculation module for the local difference matrix is ​​used to calculate the threshold of the local difference matrix and obtain the positions of m candidate target points through this threshold. The global uniqueness feature matrix acquisition module is used to calculate the global uniqueness features of each candidate target point and obtain the global uniqueness feature matrix. The specific implementation method is as follows: First, a grayscale histogram is introduced to distinguish between the target and clutter. For the k-th candidate target point, a local pixel block is first constructed on the original infrared image, centered on the coordinates of the k-th candidate target. Then, the grayscale histogram features of the local pixel block are extracted. The weighting function assigns more weight to pixels near the target center, ultimately resulting in a gray-level distance-weighted histogram of the local pixel block of the k-th candidate target point. , This refers to the total number of histogram intervals; Then, the gray-level histogram calculation is repeated for all candidate target points to obtain an m×m histogram matrix. To measure the similarity between different histograms, the Bach coefficient is used for similarity calculation, and the weighted histogram of the k-th candidate target point is calculated. The weighted histogram of the j-th candidate target point The Bartholomew's coefficient between The mathematical formula is as follows: After calculating the similarity coefficient between the weighted histogram of each candidate target point and the weighted histogram of other candidate target points, we obtain... similarity matrix Mat B In this process, the similarity value between each local pixel block and itself is reassigned to 0; a zero-matrix GU with the same size as the original image is constructed to store the global uniqueness feature values, where the element values ​​of the candidate target coordinates are calculated using the following formula: in Similarity matrix Mat B The maximum value of the elements in the k-th column. and The coordinates of the k-th candidate target point are given. After calculating the global uniqueness eigenvalues ​​of all candidate target points, the global uniqueness eigenvalue matrix GU is obtained. The fusion matrix acquisition module is used to normalize the local difference matrix and the global uniqueness matrix respectively, and then perform a dot product to obtain the fusion matrix. The infrared small target acquisition module is used to calculate the adaptive threshold of the fusion matrix, and then use this threshold to perform binarization segmentation of the fusion matrix to obtain the infrared small target.

Images