An infrared anti-interference detection method for small targets at sea

Abstract

A method for detecting infrared small targets at sea with anti-interference, which belongs to the technical field of marine target detection. The present invention solves the problems of low detection rate and high false alarm rate when using existing target detection methods due to interference in infrared sea surface images. The present invention realizes recall detection of suspected targets by calculating the dissimilarity of the grayscale distribution of the central area and its neighborhood. Then, the features of the suspected target area are further extracted. Because of the clustering characteristics of the interference at sea, a relaxed mutual k-nearest neighbor graph is introduced to distinguish the real target through global graph clustering. Through this method, the real target can be detected and the interference at sea can be filtered out. The method of the present invention can be applied to the detection of small targets at sea.

Description

Technical Field

[0001] The present invention belongs to the technical field of marine target detection, and in particular relates to an infrared marine small target anti-interference detection method. Background Art

[0002] Small infrared target detection plays a vital role in infrared image search and tracking (IRST). Small infrared targets consist of only a few pixels in an image. These small targets lack shape and texture features and are often obscured by complex background interference. Furthermore, the constantly changing background area in different infrared images can lead to varying degrees of interference. Consequently, this lack of universal characteristics complicates the detection of small infrared targets.

[0003] In recent years, extensive work has been conducted on various approaches to improve the detection of small infrared targets in single frames. Most work in this field relies on filtering in the time or frequency domain. For example, Deshpande et al. proposed maximum median and maximum mean filters, which are relatively simple in design. Contrast is a good feature for expressing the difference between foreground and background. Chen proposed the local contrast method (LCM). Wei further improved this method by proposing the multi-scale patch-based contrast measurement (MPCM), which achieved better results in target enhancement. Morphological processing has also been used for infrared small target detection. Bai improved the top-hat method, which has good performance for infrared small target detection. Using low-rank and sparse representations, Gao proposed the infrared image patch model (IPI), which transforms the detection problem into an optimization problem of recovering low-rank and sparse matrices. Dai proposed a novel reweighted infrared patch tensor model (RIPT), which uses a structured tensor and reweighting to design weights. These methods can accurately separate small targets in images with flat backgrounds. There are also some methods that use facet models for detection. Bai uses the facet model to calculate derivatives in each direction and then uses entropy features to enhance targets.

[0004] The above method performs well in common infrared scenes and can be applied to actual situations. However, in infrared sea surface images, targets are easily interfered by waves, sea surface reflections, and sea-sky lines. In particular, strong light reflections and waves on the sea surface can cause false detection or strong interference, such as Figure 1 and Figure 2 Therefore, due to the interference in the actual infrared sea surface image, the existing target detection method will have the problems of low detection rate and high false alarm rate. Summary of the Invention

[0005] The purpose of the present invention is to solve the problems of low detection rate and high false alarm rate when using existing target detection methods due to interference in infrared sea surface images, and to propose an infrared sea small target anti-interference detection method.

[0006] The technical solution adopted by the present invention to solve the above technical problems is:

[0007] A method for detecting infrared small targets at sea with anti-interference, the method specifically comprising the following steps:

[0008] Step 1: Collect infrared sea surface images and calculate the local dissimilarity measure of the area centered on each pixel in the infrared sea surface image;

[0009] Step 2: Select the pixel points of the suspected target according to the calculation results of step 1, and extract the features of the suspected target area centered on each pixel point of the suspected target;

[0010] Construct a graph G(V,E) based on the features of each suspected target area extracted;

[0011] Step 3: Based on the constructed graph G(V,E), the graph clustering method is used to divide each suspected target area into different initial classes {C1, C2, ...C R}, where C1, C2, … C R They are divided into Class 1, Class 2, ..., Class R respectively;

[0012] The number of suspected target areas in the initial class does not exceed T c The class is taken as the candidate class, the target class is filtered out from the candidate class, and the target is extracted from the suspected target area contained in the target class.

[0013] Furthermore, the specific process of step one is:

[0014] Step 1: For the pixel point (i, j) in the infrared image, use the pixel point (i, j) as an image block of size d×d The center point of the image block As the central image block, the image block As the central image patch Eight neighborhood image blocks;

[0015] Step 1 and 2: Calculate the average image block of eight neighboring image blocks

[0016] Step 13: Let X represent the central image block and average image patch The set of non-repeated gray values ​​in

[0017]

[0018] Among them, n and m represent the center image block respectively. and average image patch The number of non-repeated grayscale values ​​in , Represents the first grayscale value among n grayscale values, Represents the second grayscale value among n grayscale values, Represents the nth grayscale value among n grayscale values, Represents the first gray value among m gray values, Represents the second grayscale value among m grayscale values, Represents the mth grayscale value among m grayscale values;

[0019] Arrange the elements of set X in ascending order to form a sequence X a , and then according to the sequence X a Calculate the center image patch and average image patch The dissimilarity measure D CM (i,j);

[0020] Step 14: Use the method of step 13 to calculate the difference between each neighborhood image block and the average image block. The dissimilarity measure of the eight neighboring image blocks is then added together as the dissimilarity measure D between the average image block and the neighboring image block. NM (i,j);

[0021] Step 15: Calculate the local dissimilarity measure LDM(i,j) of the region centered at pixel point (i,j):

[0022]

[0023] Where ε is a constant;

[0024] Step 16: Similarly, the methods of steps 11 to 15 are used to obtain the local dissimilarity measure of the area centered on each pixel in the infrared sea surface image.

[0025] Furthermore, in the steps 1 and 3, according to the sequence X a Calculate the center image patch and average image patch The dissimilarity measure D CM (i,j), the specific process is:

[0026]

[0027] in, is the sequence X aThe jth grayscale value in is the sequence X a The j-1th grayscale value in the sequence X, j = 1, 2, ..., J, J + 1 is a capacity, The center image block exist The cumulative probability distribution function at , is the average image block exist The cumulative probability distribution function at .

[0028] Furthermore, the central image block exist Cumulative probability distribution function at for:

[0029]

[0030] Among them, p i′ is the grayscale value x i′ In the center image patch The probability value in the normalized grayscale histogram, grayscale value

[0031] Furthermore, the value of the constant ε is 1.

[0032] Furthermore, the specific process of step 2 is as follows:

[0033] After calculating the local dissimilarity measure of the region centered on each pixel in step 1, the local dissimilarity measure of the region is used as the local difference measure value of the corresponding central pixel point, and the LDM feature image is obtained according to the local difference measure value of each pixel point in the infrared sea surface image;

[0034] Arrange the local difference metrics of all pixels in the infrared sea surface image in descending order and select the top N pixels. LDM The pixel point corresponding to the local difference measurement value is regarded as the pixel point of the suspected target;

[0035] Take each suspected target pixel as the center and the 9×9 neighborhood of the center as the suspected target area, and then extract the LDM features of each suspected target area;

[0036] According to the extracted LDM features of each suspected target area, the LDM feature distribution histogram of each suspected target area is calculated respectively, and then the LDM feature distribution histogram of each suspected target area is subjected to amplitude Fourier transform to extract the M feature of each suspected target area respectively;

[0037] For any two suspected target regions p and q, the distance d between the suspected target regions p and q is calculated based on the extracted M features. p,q , for distance d p,q Preprocessing is performed to obtain the similarity s of the two suspected target areas p and q p,q , and then according to the similarity s p,q Calculate the normalized distance between two suspected target regions p and q

[0038] Take each suspected target area as a vertex v in the graph G(V,E), v∈V, V represents the set of vertices, and according to the normalized distance between two vertices Determine the weight between two vertices and use the weight between the two vertices as the edge in the graph G(V,E), where E represents the set.

[0039] Furthermore, the specific process of the amplitude Fourier transform is:

[0040]

[0041] Among them, M(k), 0≤k≤N-1 represents the extracted M features, m(n) is the LDM feature distribution histogram, n is the index of the eigenvalue in the LDM feature distribution histogram, n=0,1,…,N-1, N represents the number of eigenvalues ​​in the LDM feature distribution histogram, and j is the imaginary unit.

[0042] Furthermore, for any two suspected target regions p and q, the distance d between the suspected target regions p and q is calculated based on the extracted M features. p,q , for distance d p,q Preprocessing is performed to obtain the similarity s of the two suspected target areas p and q p,q , and then according to the similarity s p,q Calculate the normalized distance between two suspected target regions p and q The specific process is:

[0043]

[0044] Among them, M(k) p is the kth value in the M features of the suspected target area p, m ≥ 3, M(k) q is the kth value in the M features of the suspected target area q;

[0045] Similarity p,q for:

[0046]

[0047] Among them, σ p,q is the preprocessing parameter;

[0048]

[0049] Where NN(p,k′) is the index of the K suspected target regions closest to the suspected target region p, k′=1,2,…,K, and NN(q,k′) is the index of the K suspected target regions closest to the suspected target region q;

[0050] Normalized distance for:

[0051]

[0052] Furthermore, the normalized distance between the two vertices Determine the weight between two vertices. The specific process is:

[0053] If the vertex v corresponding to the suspected target area p p The vertex v corresponding to the suspected target area q q Between: v q Belong to v p k nearest neighbors and v p Belong to v q c×k nearest neighbors, then vertex v p With vertex v q The weight between Otherwise, vertex v p With vertex v q The weight between them is ∞, where c≥1.

[0054] Furthermore, the target class is screened out from the candidate class, and the target is extracted from the suspected target area contained in the target class; the specific process is:

[0055] 1) When the candidate class includes only one class, the candidate class is directly used as the target class;

[0056] 2) When the candidate class includes two classes, if the number of suspected target areas contained in the two classes is different, the class with fewer suspected target areas is selected as the target class; if the number of suspected target areas contained in the two classes is the same, the class with the smaller intra-class distance is selected as the target class;

[0057] 3) When the number of categories included in the candidate class is greater than 2, the target class C is filtered out from the candidate class. t The way is:

[0058]

[0059] Among them, T is the size of the index number set of candidate categories, Inter represents the shortest distance between two categories, and Cg represents the g-th category in the candidate category, g = 1, 2, ..., G, and G represents the total number of categories included in the candidate category;

[0060] Then calculate the mean value μ of the local difference metric of all pixels contained in the target class t and standard deviation σ t , according to the mean μ t and standard deviation σ t Calculate the adaptive threshold TH:

[0061] TH=μ t +k t σ t

[0062] Among them, k t is an empirical parameter;

[0063] Among all the pixels included in the target class, the pixels whose local difference metric values ​​are greater than the adaptive threshold TH are taken as target pixels.

[0064] The beneficial effects of the present invention are:

[0065] This paper proposes an anti-interference detection method for infrared small maritime targets. This method achieves recall detection of suspected targets by calculating the dissimilarity of the grayscale distribution between the central region and its neighborhood. Features of the suspected target region are then extracted. Due to the clustering characteristics of maritime interference, a relaxed mutual k-nearest neighbor graph is introduced to distinguish true targets through global graph clustering. This method can detect true targets and filter out maritime interference.

[0066] Experiments on three real-world infrared ocean datasets and one general multi-scene dataset demonstrate that the proposed method successfully detects small targets and suppresses background interference. Especially against complex backgrounds, the proposed method outperforms existing methods with high detection rates and low false alarm rates. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is infrared sea surface image 1;

[0068] Figure 2 This is the second infrared sea surface image;

[0069] Figure 3 It is a nested structure diagram of the central image block and the eight neighborhood image blocks;

[0070] Figure 4(a) is an infrared sea surface image with a small boat target marked;

[0071] Figure 4(b) is D CM image;

[0072] Figure 4(c) is DNM image;

[0073] Figure 4(d) is the result image obtained by LDM;

[0074] Figure 5(a) is sample image 1 in dataset 1;

[0075] Figure 5(b) is sample image 2 in dataset 1;

[0076] Figure 5(c) is sample image 3 in dataset 1;

[0077] Figure 5(d) is sample image 4 in dataset 1;

[0078] Figure 6(a) is sample image 1 in dataset 2;

[0079] Figure 6(b) is sample image 2 in dataset 2;

[0080] Figure 6(c) is sample image 3 in dataset 2;

[0081] Figure 6(d) is sample image 4 in dataset 2;

[0082] Figure 7(a) is sample image 1 in dataset 3;

[0083] Figure 7(b) is sample image 2 in dataset 3;

[0084] Figure 7(c) is sample image 3 in dataset 3;

[0085] Figure 7(d) is sample image 4 in dataset 3;

[0086] Figure 8 It is the ROC curve of each method on data set 1;

[0087] Figure 9 It is the ROC curve of each method on data set 2;

[0088] Figure 10 It is the ROC curve of each method on data set 3;

[0089] Figure 11(a) is an example image 1 in the general multi-scene dataset;

[0090] Figure 11(b) is an example image 2 from the general multi-scene dataset;

[0091] Figure 11(c) is example image 3 from the general multi-scene dataset;

[0092] Figure 11(d) is an example image 4 from the general multi-scene dataset;

[0093] Figure 12It is the ROC curve diagram of each method on the common multi-scene dataset. DETAILED DESCRIPTION

[0094] Specific embodiment 1: This embodiment describes an infrared small target anti-interference detection method at sea, the method specifically comprising the following steps:

[0095] Step 1: Collect infrared sea surface images and calculate the local dissimilarity measure (LDM) of the area centered at each pixel in the infrared sea surface image;

[0096] Step 2: Select the pixel points of the suspected target according to the calculation results of step 1, and extract the features of the suspected target area centered on each pixel point of the suspected target;

[0097] Construct a graph G(V,E) based on the features of each suspected target area extracted;

[0098] Step 3: Based on the constructed graph G(V,E), Peter Kontschieder's graph clustering method is used to divide each suspected target area into different initial classes {C1, C2, ... C R}, where C1, C2, … C R They are divided into Class 1, Class 2, ..., Class R respectively;

[0099] The number of suspected target areas in the initial class does not exceed T c The class is taken as the candidate class, the target class is filtered out from the candidate class, and the target is extracted from the suspected target area contained in the target class.

[0100] In actual situations, the interference sources of infrared images are various, such as waves, cumulus clouds, bright lights on the sea surface, etc. After graph clustering, suspicious targets including interference sources may also be classified into these categories. c A class with no more than T members is considered an interference class. c The smaller classes are identified as candidate classes.

[0101] The method of the present invention divides small target detection in complex infrared sea surface scenes into two stages. First, suspected target areas with local dissimilarity metrics are determined. During this process, the dissimilarity of the local area centered on each pixel is calculated, all candidate targets that meet the conditions are enhanced, and edge noise is suppressed. Subsequently, features of the suspected target area are extracted, and global graph clustering is used to distinguish the results. In this stage, false targets on the sea surface are filtered out to achieve better detection. Graph clustering is used to filter out false targets, leaving real targets. The method proposed in the present invention solves the problem of sea surface interference in infrared small target detection.

[0102] It is generally considered that a target area not exceeding 9*9 pixels is a small target area.

[0103] Specific implementation method 2: Combination Figure 3 This embodiment is different from the first embodiment in that the specific process of step 1 is as follows:

[0104] Step 1: For the pixel point (i, j) in the infrared image, use the pixel point (i, j) as an image block of size d×d The center point of the image block As the central image block, the image block As the central image patch The eight neighborhood image blocks of , each of which has a size of d×d;

[0105] Step 1 and 2: Calculate the average image block of eight neighboring image blocks (For any pixel in the average image block, the grayscale value of the pixel is the average of the grayscale values ​​of the pixels at the corresponding positions in the eight neighboring image blocks);

[0106] Step 13: Let X represent the central image block and average image patch The set of non-repeated gray values ​​in

[0107]

[0108] Among them, n and m represent the center image block respectively. and average image patch The number of non-repeated grayscale values ​​in , Represents the first grayscale value among n grayscale values, Represents the second grayscale value among n grayscale values, Represents the nth grayscale value among n grayscale values, Represents the first gray value among m gray values, Represents the second grayscale value among m grayscale values, Represents the mth gray value among m gray values;

[0109] Arrange the elements of set X in ascending order to form a sequence X a , and then according to the sequence X a Calculate the center image patch and average image patch The dissimilarity measure D CM (i,j);

[0110] The grayscale values ​​in steps 1 and 3 are the normalized grayscale values ​​of the pixels in their respective image blocks;

[0111] Step 14: Use the method of step 13 to calculate the difference between each neighborhood image block and the average image block. The dissimilarity measure of the eight neighboring image blocks is then added together as the dissimilarity measure D between the average image block and the neighboring image block. NM (i,j);

[0112] Step 15: Calculate the local dissimilarity measure LDM(i,j) of the region centered at pixel point (i,j):

[0113]

[0114] Where ε is a constant;

[0115] Step 16: Similarly, the methods of steps 11 to 15 are used to obtain the local dissimilarity measure of the area centered on each pixel in the infrared sea surface image.

[0116] LDM describes the difference in grayscale value between the central pixel and surrounding pixels within a region. It represents the characteristic value of internal rejection; larger LDMs indicate greater rejection. LDMs are calculated over the local regions of the central block and neighboring blocks.

[0117] Figure 4(a) is a typical infrared sea surface image, where the target is a small boat marked with a rectangle. The difference measure D between the center image patch and the average image patch is: CM As shown in Figure 4(b). Figure 4(c) clearly shows that D NM The image contains clutter such as the sea-skyline, wave edges, and cloud edges. Figure 4(d) shows the target enhancement and background interference suppression achieved by LDM. It can be seen that target enhancement and background interference suppression are achieved in local areas. Next, from a global perspective, graph clustering is designed to distinguish real targets from suspicious targets.

[0118] Other steps and parameters are the same as those in the first embodiment.

[0119] Specific embodiment three: This embodiment is different from specific embodiment one or two in that, in step one or three, according to sequence X a Calculate the center image patch and average image patch The dissimilarity measure D CM (i,j), the specific process is:

[0120]

[0121] in, is the sequence X a The jth grayscale value in is the sequence Xa The j-1th grayscale value in the sequence X, j = 1, 2, ..., J, J + 1 is a capacity, The center image block exist The cumulative probability distribution function at , is the average image block exist The cumulative probability distribution function at .

[0122] Other steps and parameters are the same as those in the first or second embodiment.

[0123] Specific embodiment 4: This embodiment is different from any one of specific embodiments 1 to 3 in that the central image block exist Cumulative probability distribution function at for:

[0124]

[0125] Among them, p i′ is the grayscale value x i′ In the center image patch The probability value in the normalized grayscale histogram, grayscale value

[0126] The other steps and parameters are the same as those in the first to third embodiments.

[0127] Specific implementation method five: This implementation method is different from any one of specific implementation methods one to four in that the value of the constant ε is 1.

[0128] The other steps and parameters are the same as those in the first to fourth embodiments.

[0129] Specific embodiment 6: This embodiment differs from specific embodiments 1 to 5 in that the specific process of step 2 is as follows:

[0130] After calculating the local dissimilarity measure of the region centered on each pixel in step 1, the local dissimilarity measure of the region is used as the local difference measure value of the corresponding central pixel point, and the LDM feature image is obtained according to the local difference measure value of each pixel point in the infrared sea surface image;

[0131] Arrange the local difference metrics of all pixels in the infrared sea surface image in descending order and select the top N pixels. LDM The pixel points corresponding to the local difference measurement values ​​(the value is 20 in this invention) are regarded as the pixel points of the suspected target;

[0132] Take each suspected target pixel as the center and the 9×9 neighborhood of the center as the suspected target area, and then extract the LDM features of each suspected target area;

[0133] If a suspected pixel is located in the area of ​​another suspected pixel and its LDM value is smaller, it will be suppressed by the suspected pixel with the larger LDM value. The suppressed suspected pixel will no longer be considered in the next stage;

[0134] According to the extracted LDM features of each suspected target area, the LDM feature distribution histogram of each suspected target area is calculated respectively, and then the LDM feature distribution histogram of each suspected target area is subjected to amplitude Fourier transform to extract the M feature of each suspected target area respectively;

[0135] For any two suspected target regions p and q, the distance d between the suspected target regions p and q is calculated based on the extracted M features. p,q , for distance d p,q Preprocessing is performed to obtain the similarity s of the two suspected target areas p and q p,q , and then according to the similarity s p,q Calculate the normalized distance between two suspected target regions p and q

[0136] In infrared images, the mechanism of generating interference is the same, so this type of interference is always similar to each other. These similar interference areas are easy to cluster. Therefore, the mutual k-nearest neighbor graph is introduced for global graph clustering. Each suspected target area is regarded as a vertex v in the graph G(V,E), v∈V, V represents the set of vertices, and the normalized distance between two vertices is used to cluster the two areas. Determine the weight between two vertices and use the weight between the two vertices as the edge in the graph G(V,E), where E represents the set.

[0137] The other steps and parameters are the same as those in the first to fifth embodiments.

[0138] Specific embodiment seven: This embodiment differs from any one of specific embodiments one to six in that the specific process of the amplitude Fourier transform is as follows:

[0139]

[0140] Among them, M(k), 0≤k≤N-1 represents the extracted M features, m(n) is the LDM feature distribution histogram, n is the index of the eigenvalue in the LDM feature distribution histogram, n=0,1,…,N-1, N represents the number of eigenvalues ​​in the LDM feature distribution histogram, and j is the imaginary unit.

[0141] The other steps and parameters are the same as those in the first to sixth embodiments.

[0142] Specific embodiment eight: This embodiment differs from any one of specific embodiments one to seven in that, for any two suspected target regions p and q, the distance d between the suspected target regions p and q is calculated based on the extracted M features. p,q , for distance d p,q Preprocessing is performed to obtain the similarity s of the two suspected target areas p and q p,q , and then according to the similarity s p,q Calculate the normalized distance between two suspected target regions p and q The specific process is:

[0143]

[0144] Among them, M(k) p is the kth value in the M features of the suspected target region p, m ≥ 3 (the minimum value of m is 3, which is used to achieve the most comprehensive measurement between suspected target regions), M(k) q is the kth value in the M features of the suspected target area q;

[0145] s p,q The smaller the value, the less likely p and q are to belong to the same category. p,q for:

[0146]

[0147] Among them, σ p,q is the preprocessing parameter (it aggregates suspected targets belonging to the same class as much as possible and separates suspected targets belonging to different classes);

[0148]

[0149] Where NN(p,k′) is the index of the K suspected target regions closest to the suspected target region p, k′=1,2,…,K, and NN(q,k′) is the index of the K suspected target regions closest to the suspected target region q;

[0150] Normalized distance for:

[0151]

[0152] The other steps and parameters are the same as those in the first to seventh embodiments.

[0153] Specific embodiment 9: This embodiment differs from any one of specific embodiments 1 to 8 in that the normalized distance between two vertices is used. Determine the weight between two vertices. The specific process is:

[0154] If the vertex v corresponding to the suspected target area p p The vertex v corresponding to the suspected target area q q Between: v q Belong to v p k nearest neighbors and v p Belong to v q c×k nearest neighbors, then vertex v p With vertex v q The weight between Otherwise, vertex v p With vertex v q The weight between them is ∞, where c≥1.

[0155] The other steps and parameters are the same as those in Specific Embodiments 1 to 8.

[0156] Specific embodiment 10: This embodiment differs from specific embodiments 1 to 9 in that the target class is screened out from the candidate class, and the target is extracted from the suspected target area contained in the target class; the specific process is as follows:

[0157] 1) When the candidate class includes only one class, the candidate class is directly used as the target class;

[0158] 2) When the candidate class includes two classes, if the number of suspected target areas contained in the two classes is different, the class with fewer suspected target areas is selected as the target class; if the number of suspected target areas contained in the two classes is the same, the class with the smaller intra-class distance is selected as the target class;

[0159] 3) When the number of categories included in the candidate class is greater than 2, the target class C is filtered out from the candidate class. t The way is:

[0160]

[0161] Among them, T is the size of the index number set of candidate categories, Inter represents the shortest distance between two categories, and C g represents the g-th category in the candidate category, g = 1, 2, ..., G, and G represents the total number of categories included in the candidate category;

[0162] Then calculate the mean value μ of the local difference metric of all pixels contained in the target class t and standard deviation σ t , according to the mean μ t and standard deviation σ t Calculate the adaptive threshold TH:

[0163] TH=μ t +kt σ t

[0164] Among them, k t is an empirical parameter;

[0165] Among all the pixels included in the target class, the pixels whose local difference metric values ​​are greater than the adaptive threshold TH are taken as target pixels.

[0166] The other steps and parameters are the same as those in Specific Embodiments 1 to 9.

[0167] Experimental part

[0168] To demonstrate the superiority and robustness of the proposed method, we tested it using a dataset of infrared small targets at sea, consisting of three different scenes. Dataset 1 consists of 204 images with a resolution of 740 × 470 pixels, and the target is a small boat near the sea-skyline. Dataset 2 consists of 98 images with a resolution of 740 × 340 pixels, and the target is a floating sphere on the sea surface. Dataset 3 consists of 98 images with a resolution of 740 × 470 pixels, and the target is a hot object located in a mountain above the coastline.

[0169] Several infrared small target detection methods were compared, including the morphological method Top-Hat, the contrast-based method MPCM, the sparse-based method IPI, the local intensity and gradient-based methods LIG and FAMSIS, the facet kernel, and the RW-based method FKRW. The background suppression factor (BSF) is commonly used to evaluate the performance of small target detection methods. It is defined as follows:

[0170]

[0171] σ in and σ out The BSF is used to measure the residual degree of background clutter and noise in the result image. In infrared sea surface images, background clutter is widely distributed and mainly caused by waves. The complex background in the images of the three datasets produces a high standard deviation σ. in A good detection method requires not only the correct detection of small targets, but also strong background interference suppression capabilities. A higher BSF indicates better performance in target detection and background suppression. Therefore, the resulting image of a good detection method always has a lower standard deviation σ out We add a coefficient λ to the denominator of BSF to avoid the denominator being 0. In the present invention, λ is set to 0.1.

[0172] We use the average error (AE) to measure the performance of enhancing the actual target and suppressing background interference, as shown in formula (2).

[0173]

[0174] In order to facilitate statistics and analysis, the detection result images of all methods are Normalized to [0,255] using formula (3). mak Represents a mask image with the target area filled with 255 and the remaining pixels set to 0. pixs Represents the number of pixels in the image.

[0175]

[0176] In addition, using the single-frame detection probability P d and false alarm rate F a , defined as follows:

[0177]

[0178]

[0179] At the same time, if two conditions are met: if there is an overlapping area between the detected target and the actual target, and the Chebyshev distance between the center of mass of the detected target area and the actual target cannot exceed the threshold (three pixels), then the target is considered to be a true detection. d and false alarm rate F a A receiver operating characteristic (ROC) curve can be drawn to describe the P d and F a The dynamic relationship between them. The AUC value (area under the ROC curve) of each method can be calculated based on the curve (the larger the value, the better the performance).

[0180] Infrared sea surface dataset

[0181] We used BSF, AE, and ROC curves to evaluate the detection performance of the proposed method and other comparison methods. In Dataset 1, the ocean waves are the primary distractor, and the small target is clearly visible. In Dataset 2, the high-brightness rock wall is the primary background distractor, and the small target is relatively dim. In Dataset 3, the mountain and ocean waves are the primary background distractors, and the small target is relatively dim. Figure 5(a) to Figure 5(d) Some examples from dataset 1 are shown, Figure 6(a) to Figure 6(d) Some examples from dataset 2 are shown, Figure 7(a) to Figure 7(d) Some examples from Dataset 3 are shown.

[0182] Table 1 shows and From the results of the method on three datasets, it can be seen that the method proposed in the present invention outperforms other methods in both BSF and AE.

[0183] Table 1 Evaluation criteria for the infrared sea surface dataset

[0184]

[0185] Figure 8 The ROC curves of the methods in Dataset 1 are shown. As can be seen, as the detection threshold increases, most methods maintain a high detection probability. The method proposed in this paper maintains the highest accuracy.

[0186] Figure 9 The ROC curves for each method in Dataset 2 are shown. The method proposed in this paper still achieved the best results among all methods, namely the highest AUC. Furthermore, compared with Dataset 1, the performance of all methods (except the method proposed in this paper) deteriorated significantly. This phenomenon may be due to the strong interference from the high-brightness rock wall, which led to a high number of false detections. The method proposed in this paper can overcome this interference.

[0187] Figure 10 The ROC curves for each method in Dataset 3 are shown. The proposed method again performs best among all methods, with the highest AUC. All methods except the proposed method perform worse than those in Dataset 2. This phenomenon is caused by the darker true target, which is difficult to detect due to interference. For the proposed method, the ocean waves present little interference.

[0188] Infrared general multi-scene dataset

[0189] We also conduct experimental comparisons on a total of 126 collected common multi-scene infrared images. Figure 11(a) to Figure 11(d) Some example images from a common multi-scene dataset are shown. Sometimes, disturbances in normal situations also have clustering characteristics similar to waves, so the method of the present invention can still be used. Figure 12 As can be seen from Table 2, the method of the present invention still has some advantages, such as ROC curve, BSF and AE, which shows that the method of the present invention can also be used in common multi-scenario situations.

[0190] Table 2 Evaluation criteria for infrared general multi-scene dataset

[0191]

[0192] This paper proposes a global graph clustering-based infrared small target detection algorithm, called the interference-resistant local dissimilarity measure based on global graph clustering (LDMGGC). The main idea of ​​this method is to exploit the uniqueness of the true target in its grayscale distribution and the clustering characteristics of different types of sea surface interference.

[0193] The experimental results of dataset 1 show that when BSF and AE are used as performance indicators, the performance of the LDMGGC method proposed in the present invention is very good, and a smaller comparative advantage is obtained under the ROC curve (because there are more cluster interferences and easier targets). The results of datasets 2 and 3 show that LDMGGC has a greater comparative advantage when the ROC curve is used, and a smaller comparative advantage is obtained when BSF and AE are considered (because there are more cluster interferences and darker targets). The results on the infrared ordinary multi-scene dataset show that LDMGGC has a smaller comparative advantage on the ROC curve, and BSF and AE also have a smaller comparative advantage due to less cluster interference. However, it can be seen that even with a weaker advantage, LDMGGC is still useful in ordinary infrared multi-scene situations.

[0194] The above examples are merely illustrative of the calculation model and process of the present invention and are not intended to limit the embodiments of the present invention. Persons skilled in the art will readily appreciate that other variations or modifications based on the above description are possible. This list of embodiments is not exhaustive; however, any obvious variations or modifications derived from the technical solution of the present invention remain within the scope of protection of the present invention.

Claims

1. A method for detecting infrared small targets at sea with anti-interference, characterized in that: The method specifically comprises the following steps: Step 1: Collect infrared sea surface images and calculate the local dissimilarity measure of the area centered on each pixel in the infrared sea surface image; The specific process of step one is: Step 1: For the pixel points in the infrared image , in pixels As the size of Image blocks The center point of the image block As the central image block, the image block , As the central image patch Eight neighborhood image blocks; Step 1 and 2: Calculate the average image block of eight neighboring image blocks ; Step 13: Represents the central image patch and average image patch The set of non-repeated gray values ​​in in, and Represents the center image block and average image patch The number of non-repeated grayscale values ​​in , express The first gray value among the gray values, express The second gray value among the gray values, express The gray value Gray values, express The first gray value among the gray values, express The second gray value among the gray values, express The gray value Grayscale values; will be collected The elements in are arranged in ascending order to form a sequence , and then according to the sequence Calculate the center image patch and average image patch Dissimilarity measure ; in, is a sequence The Gray values, is a sequence The Gray values, , is a sequence capacity, The center image block exist The cumulative probability distribution function at , is the average image block exist Cumulative probability distribution function at ; Step 14: Use the method of step 13 to calculate the difference between each neighborhood image block and the average image block. The dissimilarity measure of the eight neighboring image blocks is then added together as the dissimilarity measure between the average image block and the neighboring image blocks. ; Step 15: Calculate the pixel The local dissimilarity measure of the region centered on : in, is a constant; Step 16: Similarly, the methods of steps 11 to 15 are used to respectively obtain the local dissimilarity measure of the area centered at each pixel in the infrared sea surface image; Step 2: Select the pixel points of the suspected target according to the calculation results of step 1, and extract the features of the suspected target area centered on each pixel point of the suspected target; Construct a map based on the extracted features of each suspected target area ; Step 3: Based on the constructed graph , using graph clustering method to divide each suspected target area into different initial classes ,in, They are divided into the first category, the second category, ..., kind; The number of suspected target areas in the initial class does not exceed The class is taken as the candidate class, the target class is filtered out from the candidate class, and the target is extracted from the suspected target area contained in the target class.

2. The infrared small target anti-interference detection method according to claim 1 is characterized in that: The central image block exist Cumulative probability distribution function at for: in, is the grayscale value In the center image patch The probability value in the normalized grayscale histogram, grayscale value .

3. The infrared small target anti-interference detection method according to claim 2 is characterized in that: The constant The value of is 1.

4. The infrared small target anti-interference detection method according to claim 3 is characterized in that: The specific process of step 2 is as follows: After calculating the local dissimilarity measure of the region centered on each pixel in step 1, the local dissimilarity measure of the region is used as the local difference measure value of the corresponding central pixel point, and the LDM feature image is obtained according to the local difference measure value of each pixel point in the infrared sea surface image; Arrange the local difference metrics of all pixels in the infrared sea surface image in descending order and select the top The pixel point corresponding to the local difference measurement value is regarded as the pixel point of the suspected target; Take each suspected target pixel as the center and the 9×9 neighborhood of the center as the suspected target area, and then extract the LDM features of each suspected target area; According to the extracted LDM features of each suspected target area, the LDM of each suspected target area is calculated respectively. Feature distribution histogram, and then each suspected target area The feature distribution histogram is subjected to amplitude Fourier transform to extract each suspected target area. feature; For any two suspected target areas and , according to the extracted Feature calculation of suspected target area and The distance between , for distance Preprocessing to obtain two suspected target areas and Similarity , and then according to the similarity Calculate two suspected target areas and The normalized distance between ; Each suspected target area is used as a A vertex in , , Represents a set of vertices, based on the normalized distance between two vertices Determine the weight between two vertices and use the weight between the two vertices as the graph The edge of A collection of representatives.

5. The infrared small target anti-interference detection method according to claim 4 is characterized in that: The specific process of the amplitude Fourier transform is: in, , Represents the extracted feature, yes Feature distribution histogram, yes The index of the eigenvalue in the feature distribution histogram, , represent The number of eigenvalues ​​in the feature distribution histogram, Is an imaginary unit.

6. The infrared small target anti-interference detection method according to claim 5 is characterized in that: For any two suspected target areas and , according to the extracted Feature calculation of suspected target area and The distance between , for distance Preprocessing to obtain two suspected target areas and Similarity , and then according to the similarity Calculate two suspected target areas and The normalized distance between The specific process is as follows: in, Suspected target area of The first of the features values, , Suspected target area of The first of the features values; Similarity for: in, is the preprocessing parameter; in, Is the suspected target area The closest The index of the suspected target area, , Is the suspected target area The closest The index of the suspected target area; Normalized distance for: 。 7. The infrared small target anti-interference detection method according to claim 6 is characterized in that: The normalized distance between two vertices Determine the weight between two vertices. The specific process is: If the target area is suspected The corresponding vertex and suspected target areas The corresponding vertex Satisfy between: belong of Neighbors and belong of Nearest neighbor, then the vertex With vertex The weight between , otherwise the vertex With vertex The weight between ,in, .

8. The infrared small target anti-interference detection method according to claim 7 is characterized in that: The target class is screened out from the candidate class, and the target is extracted from the suspected target area contained in the target class; the specific process is: 1) When there is only one class in the candidate class, the candidate class is directly used as the target class; 2) When the candidate class includes two classes, if the number of suspected target areas contained in the two classes is different, the class with fewer suspected target areas is selected as the target class; if the number of suspected target areas contained in the two classes is the same, the class with the smaller intra-class distance is selected as the target class; 3) When the number of categories included in the candidate class is greater than 2, the target class is filtered out from the candidate class. The way is: in, is the size of the set of index numbers of candidate categories, represents the shortest distance between two classes, Representative candidate category kind, , represents the total number of categories included in the candidate category; Then calculate the mean of the local difference metrics of all pixels contained in the target class and standard deviation , according to the mean and standard deviation Calculating adaptive thresholds : in, is an empirical parameter; Among all pixels contained in the target class, the local difference metric value is greater than the adaptive threshold. The pixel point is taken as the target pixel point.

Images