Underwater small target detection method and system based on sonar images under complex background

Abstract

The present invention discloses a method and system for detecting underwater small targets in sonar images under complex backgrounds. The method comprises: denoising the collected sonar images using a multi-level median filtering method; performing complex terrain masking on the denoised sonar images to extract highlight texture maps and shadow texture maps of the terrain; performing high-order moment image segmentation on the denoised sonar images, and combining the complex terrain masking results to perform background suppression; clustering the processed connected domains using three different-sized detection windows to associate split targets and suppress ridge terrain; extracting highlight areas and shadow areas, calculating highlight pixel density and shadow pixel density respectively, and obtaining a detection score through pixel density weighted calculation to detect small targets. The method is not restricted by seabed terrain and adaptively adjusts the threshold according to the required false alarm rate and actual acoustic image intensity. It has the characteristics of high accuracy, low false alarm rate, and strong adaptability.

Description

Technical Field

[0001] The present invention relates to the field of underwater target detection, and in particular to a method and system for detecting small underwater targets in sonar images under complex backgrounds. Background Art

[0002] Research on underwater target detection technology is a challenging issue in the field of underwater acoustics, and is particularly critical for applications such as underwater target detection and sunken ship and object recovery. Compared to other underwater detection equipment, high-resolution imaging sonar offers significant advantages such as high efficiency, high resolution, and long detection range, making it the primary means of detecting and identifying small, stationary underwater targets.

[0003] However, current imaging sonar underwater target detection suffers from high false alarm rates and low accuracy. The main reasons for this problem are: 1) Complex topography. In addition to flat mud and sand, the seabed also features ridges, slopes, furrows, and cracks, and multiple landforms can exist in the same area. 2) Low acoustic image resolution and high noise. Because the wavelength of the sound waves is close to the target size, sonar images have low resolution and limited texture information. Due to multipath, reverberation, and the randomness of the medium itself, as well as variations in parameters such as reflection and scattering, sonar images contain more noise than optical images. Motion errors can cause highlight areas to become elongated, significantly deviating from their true size, and shadows in the acoustic image are weak or absent. 3) Few labeled target samples are available. Due to high acquisition costs, military confidentiality, and the complex underwater environment, obtaining raw acoustic image target samples is difficult, and some targets are even completely unavailable. Furthermore, target labeling is difficult, even for cooperative targets. This results in a small number of target samples and an uneven distribution between target classes.

[0004] Typically, classic underwater target detection algorithms based on acoustic images use carefully designed sliding windows to search for targets over a wide area, then leverage statistical features of highlight and shadow shape contours for detection. However, when sonar images are blurry and the target size in the acoustic image differs significantly from the actual size, missed detections are common. In recent years, deep learning, with its superior automatic feature extraction capabilities, has significantly improved target detection accuracy, and some researchers have applied deep learning techniques to underwater target detection based on sonar images. However, the superior performance of deep learning relies on complex network structures and large-scale annotated datasets. When labeled sonar image samples are scarce, the discrepancy between the actual scene and the training samples can lead to mismatches, resulting in decreased accuracy and increased false alarm rates. Summary of the Invention

[0005] The purpose of the present invention is to overcome the defects of the prior art and propose a method and system for detecting underwater small targets in sonar images under complex backgrounds.

[0006] To achieve the above object, the present invention proposes a method for detecting small underwater targets in a sonar image under a complex background, the method comprising:

[0007] Step 1) adopting a multi-level median filtering method to denoise the collected sonar image;

[0008] Step 2) performing complex terrain masking processing on the sonar image after denoising in step 1) to extract the highlight texture map and shadow texture map of the terrain;

[0009] Step 3) performing high-order moment image segmentation on the sonar image after denoising in step 1), and combining the complex terrain mask processing result in step 2) to perform background suppression;

[0010] Step 4) clustering the connected domains processed in step 3) using three different sizes of detection windows to achieve association of split targets and suppression of ridge landforms;

[0011] Step 5) Extract highlight areas and shadow areas, calculate highlight pixel density and shadow pixel density respectively, and obtain detection scores through pixel density weighted calculation to achieve small target detection.

[0012] As an improvement to the above method, step 1) includes:

[0013] The sonar image is denoised by traversing each pixel of the image and calculating the median v of each pixel I(i,j) in the azimuth and range directions respectively. The pixel median u of the corner point where I(i,j) is located is then calculated. Finally, the median of I(i,j), v and u is used as the value of the pixel. Multi-level median processing is used to suppress noise and retain the edge contour of the target.

[0014] As an improvement to the above method, step 2) includes:

[0015] Step 2-1) Using the integral graph method, background accumulation is performed on the sonar image after denoising in step 1) to obtain a cumulative graph B(i, j);

[0016] Step 2-2) Using the statistical characteristics of the acoustic seabed topography that obeys the Rayleigh distribution, according to the required false alarm rate P fa Calculate the threshold V T :

[0017]

[0018] Where σ is the standard deviation of the Rayleigh distribution, and the threshold coefficient ρ is obtained from the following formula:

[0019]

[0020] Step 2-3) Using the cumulative image from step 2-1), perform large-scale sliding window processing and calculate the average value M(k) of the acoustic image in the k-th sliding window; when the cumulative image B(i,j)>M(k)·ρ, the pixel belongs to the highlight texture image H B ; When B(i,j)>M(k) / ρ, the pixel belongs to the shadow texture map S B ;

[0021] Step 2-4) Use morphological operators to fill holes in the connected domains of the highlight texture map and the shadow texture map respectively; then remove the connected domains whose area is smaller than the preset area value S. max , the length or width is less than the preset length value L max The small target connected domain is obtained and the edges of the connected domain are smoothed.

[0022] As an improvement to the above method, step 3) includes:

[0023] Step 3-1) performing a power exponential transformation on the sonar image de-noised in step 1);

[0024] Step 3-2) Determine the neighborhood window Ω of any pixel (i, j) in the image G after power exponential transformation i,j , calculate the kurtosis value of the pixel HOG(i,j) according to the mean value in the neighborhood window:

[0025]

[0026] Where S i,j Represents the number of pixels in the neighborhood window, (p,q) is the neighborhood window Ω i,j Pixel coordinates within ;

[0027] Step 3-3) Calculate the threshold coefficient ρ based on the false alarm rate, and achieve highlight and shadow segmentation by judging HOG(i,j):

[0028] When the kurtosis value HOG(i,j) is less than ρ·η(p,q), the pixel belongs to the highlight area H T ;

[0029] When the kurtosis value HOG(i,j) is greater than η(p,q) / ρ, the pixel belongs to the shadow area S T ;

[0030] Step 3-4) Use morphological operators to remove the areas larger than S max and the area is smaller than S min area and perform edge smoothing;

[0031] Step 3-5) Through the highlighted area H T , and the highlight texture image H obtained in step 2) B and shadow texture map SB Differential processing is performed separately to achieve background suppression.

[0032] As an improvement to the above method, step 4) includes:

[0033] Step 4-1) Count the highlighted area graph H T Each connected domain in the network is found, and the area, center position and circumscribed rectangular frame of the connected domain are calculated.

[0034] Step 4-2) Jointly count the pixel densities of highlights and shadows, and use a small detection window to determine whether adjacent connected domains are associated;

[0035] Step 4-3) Using the middle detection window, determine whether the connected domains in the high-density area are associated;

[0036] Step 4-4) By determining the strength and direction angle of the connected domain of the high-density area within the large detection window, the areas with distant distances including ridges and cables and scattered connected domains are associated.

[0037] As an improvement to the above method, step 4-2) includes:

[0038] Sort by the center position of the connected domains, then expand the upper, lower, left, and right pixels of the circumscribed rectangle of each connected domain. Determine whether it intersects with adjacent connected domains based on the border position relationship, and thus determine whether it is associated.

[0039] As an improvement to the above method, step 4-3) includes:

[0040] Step 4-3-1) Highlight the area H T Each connected region in is considered as a point P m , the position of the point is the center of the connected domain, m represents the sequence number of the point, and the statistics are based on point P m is the center and the side length is L med The number of points in the square area, when the number of points is greater than the set high-density area threshold, the area is a high-density area;

[0041] Step 4-3-2) traverse each high-density area and determine whether adjacent high-density areas intersect. If the areas intersect, merge the intersecting areas to finally obtain a set of high-density areas;

[0042] Step 4-3-3) From the highlighted area H T Remove the connected domains in high-density areas.

[0043] As an improvement to the above method, step 4-4) includes:

[0044] Step 4-4-1) Highlight the area H TThe remaining connected domain is regarded as point P m , statistics are based on the side length L large The number of points in the square area is used to determine whether the area is a high-density area according to the methods of 4-3-1) and 4-3-2), and the high-density areas are merged and sorted according to the center coordinates of the merged high-density areas;

[0045] Step 4-4-2) Calculate the vector connected domain P in the high-density area m and point P n When the distance is less than the threshold D th When , go to step 4-4-3); otherwise, loop to the next point; where n represents the serial number of the adjacent point;

[0046] Step 4-4-3) When point P m Not fused with other points, go to step 4-4-4) to judge the intensity ratio; when point P m It has been fused with other points, and the number of fused points is less than 3. Calculate the direction angle of all point differences. Est ; When the number of fusion points is greater than or equal to 3, only the last fused 3 points are used to calculate the direction angle angle Est , then calculate the point P to be integrated n Angle after adding Pre , when d angle =|angle Pre -angle Est |Less than the angle difference threshold A th If yes, go to step 4-4-4) for further judgment; otherwise, loop to the next point;

[0047] Step 4-4-4) Calculate the two adjacent connected regions P m and P n The corresponding average intensity in the original image, and calculate its intensity ratio R meanI (m,n):

[0048]

[0049] When the intensity ratio R meanI (m,n) is greater than the threshold R th When P m and P n Fusion; among them, meanI m and meanI n Represents point P respectively m and P n The connected domain referred to corresponds to the average intensity value in the original image;

[0050] Step 4-4-5) loops for the next point and goes to step 4-4-2) until all points are traversed and goes to step 4-4-6);

[0051] Step 4-4-6) Take the center of the last fused point as the center of the fused point, calculate the union of the fused points, and use the union as the fused area;

[0052] Step 4-4-7) Determine whether each connected domain after fusion intersects and further fuse the clusters.

[0053] As an improvement to the above method, step 5) includes:

[0054] Step 5-1) Remove the areas that are too large and too small from the connected domain, and intercept the highlight area, shadow area and background area in the original image where the remaining connected domain is located, and calculate the average value B of the background area mean ;

[0055] Step 5-2) Calculate the pixel density of the highlight area and the shadow area;

[0056] Step 5-3) Based on the above pixel density, a weighted average method is used to calculate the detection score f of the connected domain. c ; Compare with the set value to complete small target detection.

[0057] The sonar image denoising processing module is used to perform denoising on the collected sonar images using a multi-stage median filtering method;

[0058] The complex terrain mask processing module is used to perform complex terrain mask processing on the denoised sonar image to extract the highlight texture map and shadow texture map of the terrain;

[0059] The high-order moment image segmentation module is used to perform high-order moment image segmentation on the denoised sonar image and perform background suppression in combination with the results of the complex terrain mask processing module;

[0060] The connected region clustering module is used to cluster the connected regions processed by the high-order moment image segmentation module using three different-sized detection windows, so as to achieve the association of split targets and the suppression of ridge landforms;

[0061] The small target detection module is used to extract highlight areas and shadow areas, calculate highlight pixel density and shadow pixel density respectively, and obtain a detection score through pixel density weighted calculation to achieve small target detection.

[0062] Compared with the prior art, the advantages of the present invention are:

[0063] 1. The present invention adopts an unsupervised target detection method, which is not limited by the number of samples, does not require sample training, and does not rely on the shape information of the target. It uses the acoustic reflection intensity information, adopts a hierarchical architecture, and realizes adaptive detection of underwater small targets through acoustic image statistical analysis, effectively reducing false alarms and improving target detection rate;

[0064] 2. The present invention has an adaptive threshold value, which can adaptively adjust the threshold value of image segmentation and target detection according to the given false alarm rate and acoustic image statistical distribution and the background intensity of the sonar image;

[0065] 3. The present invention is suitable for detecting small targets in complex seabed landforms including ridges, slopes, furrows, cracks, reefs, etc. By processing complex landform masks and clustering connected regions, it effectively suppresses false alarms caused by landforms and has strong adaptability.

[0066] 4. The present invention adopts an unsupervised statistical analysis method that does not require training samples and does not require detection model training, thus avoiding the problems of false alarms and missed detections caused by mismatches between training samples and actual scenes. It is suitable for acoustic image target detection and recognition of small samples;

[0067] 5. The present invention does not rely on the true shape characteristics of the target, but only uses the intensity characteristics and rough length and width information of the target to achieve small target detection. Even when the shape of the acoustic image changes significantly due to motion errors, or the shadow becomes weaker or disappears, the target can still be detected more accurately without being limited by the quality of the sonar image.

[0068] 6. The present invention mainly focuses on the detection of small, stationary underwater targets. However, through connected region association clustering, it also realizes pipeline detection and salvage of sunken ships and objects. At the same time, the present invention is also suitable for other underwater target detection such as archaeology. The scope of application is not limited to small target detection, and it has a wide range of applicable scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 Flowchart of the method for detecting small underwater targets in sonar images under complex backgrounds according to the present invention;

[0070] Figure 2 This is a schematic diagram of the sound image denoising process;

[0071] Figure 3(a) and Figure 3(b) are schematic diagrams of background accumulation, where Figure 3(a) is a schematic diagram of integral image calculation, and Figure 3(b) is a schematic diagram of background accumulation area;

[0072] Figure 4 This is the flow chart of high-order moment image segmentation;

[0073] Figure 5 The highlighted shadow position of the acoustic image of a small underwater target;

[0074] Figure 6(a)-Figure 6(d)The results of large detection window processing are shown in Figure 6(a) , which is the original sound image, Figure 6(b) is the result after segmentation, Figure 6(c) is the high-density area calculation, and Figure 6(d) is the large window association result.

[0075] Figure 7(a)-Figure 7(c) The results of large-scale acoustic image target detection are shown in Figure 7(a). Figure 7(b) shows the detected connected domain of suspected targets, and Figure 7(c) shows the score of large-scale acoustic image target detection. DETAILED DESCRIPTION

[0076] The present invention discloses a method for detecting small underwater targets in sonar images under complex backgrounds. The method comprises the following steps: 1) denoising the collected sonar images using a multi-level median filter method; 2) accumulating the denoised sonar images using integral graphs, and using the statistical characteristics of the sonar images following the Rayleigh distribution to calculate an adaptive threshold based on the false alarm rate to extract highlight and shadow texture images of the seafloor; 3) using high-order moments to highlight highlight and shadow features, and segmenting the highlights and shadows of the underwater small targets based on the threshold coefficient calculated from the false alarm rate. The method also combines the results of complex terrain masking to perform background suppression; 4) performing association clustering on the processed areas using three detection windows (large, medium, and small), effectively associating broken targets and suppressing false alarms using different clustering criteria; and 5) combining the detection density of highlights and shadows to discriminate targets and calculate a detection score. This target detection method is not restricted by seafloor topography and adaptively adjusts the threshold based on the required false alarm rate and the actual sonar image intensity. It features high accuracy, low false alarm rate, and strong adaptability.

[0077] This method uses the acoustic reflection intensity information, adopts a hierarchical architecture, and realizes adaptive underwater small target detection through acoustic image statistical analysis. The technical solution adopted by the present invention is:

[0078] 1. Sonar image denoising processing module

[0079] Smoothing the acoustic image can suppress thermal noise in the receiver's internal circuits. To better suppress noise and preserve small target information, a multi-stage median filtering method is used. First, the median v of the azimuth and range dimensions of the image pixel I(i,j) is calculated. Then, the median u of the pixel at the corner point where I(i,j) is located is calculated. Finally, the median of I(i,j), v, and u is calculated as the pixel value.

[0080] 2. Complex terrain mask processing module

[0081] In underwater target detection, complex terrain can easily cause a large number of false alarms. Therefore, before target detection, it is necessary to extract the background texture features of the acoustic image, specifically including:

[0082] 1) Use the integral image method for background accumulation. The integral image can be understood as a graph where the value of any point on the graph is the sum of the grayscale values ​​of all pixels within the region between the upper left corner of the grayscale image and the current point. To avoid being affected by the target, a segmentation window template consisting of three equal-sized strips is used, as shown in the following formula:

[0083]

[0084] Where, image pixel I(i,j), orientation τ j and distance to τ i The length is 3m, a b is the sum of the pixels included.

[0085] 2) Calculate the threshold value based on the required false alarm rate based on the statistical characteristics of the acoustic image seabed topography. The acoustic image seabed topography follows the Rayleigh distribution, and its probability density is shown as follows:

[0086]

[0087] Where σ is the standard deviation of the Rayleigh distribution, and its mean is The variance is According to the false alarm rate P fa , the threshold V is obtained by the above formula T :

[0088]

[0089] The threshold coefficient When P fa When =0.1, ρ=1.7122.

[0090] 3) Segmentation of acoustic image texture features. Sliding window processing is used for large acoustic images. Each window is 3m long and wide. The average acoustic image value M(k) (k=1,…,K) in each sliding window is calculated. When the cumulative image B(i,j)>M(k)·ρ, the pixel belongs to the highlight texture image H B ; When B(i,j)>M(k) / ρ, the pixel is the shadow texture S B .

[0091] 4) Morphological processing. Use morphological operators to fill holes in the connected domains of the highlight texture map and the shadow texture map respectively; then, remove the connected domains with an area smaller than S. max , length or width is less than L max Small target connected area, this paper sets S according to the rough size of the target max =10000 pixels, L max = 100 pixels; finally, edge smoothing is performed on the connected domain.

[0092] 3. High-order moment image segmentation module

[0093] Since small underwater man-made targets are mostly made of metal, they produce strong reflected echoes. When sound waves strike the seabed, they primarily reflect and scatter off the surface. Underwater targets appear as brighter areas on the acoustic image, casting shadows as they move away from the sonar. However, due to carrier motion errors, the target's shape can differ significantly from its actual shape. Therefore, target detection primarily utilizes the intensity information of acoustic image targets, specifically including:

[0094] 1) Image power exponential transformation, as shown below:

[0095] G=I(i,j) α

[0096] In the formula, α is 0.35;

[0097] 2) Calculation of kurtosis. Sliding window processing is adopted. The length and width of each window Ω are 0.5m×0.5m. The method for calculating the kurtosis of pixel G(i,j) is as follows:

[0098]

[0099] Where S i,j Represents the number of pixels in the neighborhood window, (p,q) is the neighborhood window Ω i,j Pixel coordinates within;.

[0100] 3) Segmentation of highlight and shadow areas. Calculate the threshold coefficient ρ based on the false alarm rate and perform highlight and shadow segmentation:

[0101]

[0102]

[0103] 4) Morphological processing: Use morphological operators to remove areas that are too small or too large, and perform edge smoothing.

[0104] 5) Background suppression: traverse H T and H B Pixels, when H B >0, H T =0; in this way, the shadow texture is suppressed, that is, S B >0, H T =0.

[0105] 4. Connected Region Clustering Module

[0106] Due to the varying heights of objects and terrain, as well as the impact of the angle of incidence, acoustic reflection intensities vary. Image segmentation can split objects and terrain into multiple connected domains. To reduce missed detections and false alarms, we use three detection windows: large, medium, and small.

[0107] 1) Statistics H T Each connected domain in the network is found, and the area, center position and circumscribed rectangular frame of the connected domain are calculated.

[0108] 2) Small detection window association. First, sort the connected domains according to their center positions. Then, expand the connected domain's circumscribed rectangle by S1 pixels above, below, left, and right (S1 is set to 20 pixels in this paper). Determine whether they intersect based on the border position relationship.

[0109] 3) Middle detection window association. This is mainly to associate and suppress the residual areas with close distance and high density in the background. Therefore, the processing is carried out according to the clustering of connected domains:

[0110] a) Count high-density areas. First, map the highlighted area H T Each connected region in is considered as a point P m (m=1,…,N), the position of the point is the center of the connected domain. Statistics are taken at any point P m is the center and the side length is L med =Number of points in a square area of ​​25m m , when N m >N th When , the area is a high-density area, N th is the high-density area threshold;

[0111] b) Merge high-density areas. Traverse each high-density area and determine whether adjacent high-density areas intersect. If the areas intersect, merge the intersecting areas and finally obtain a set of high-density areas C. h , 0 <h<N c ;

[0112] c) From H T Remove the high-density area C in the figure h connected domain.

[0113] 4) Large detection window association: For areas with long distances, such as ridges and cables, where the connected domains after segmentation are not concentrated but the intensity and slope are similar, association is performed.

[0114] a) Count high-density areas. Count any point P m is the center and the side length is L large =Number of points in a 100-pixel square area m’ When N m’ >N th When , the area is a high-density area. Then traverse each high-density area and merge the high-density areas. Finally, sort the high-density areas according to their center coordinates;

[0115] b) High-density area P at adjacent positionsm and P n , calculate the distance d between the two ij , when the distance is less than d ij <D th , enter step c). Otherwise, loop to the next point;

[0116] c) Direction angle discrimination. When the connected component P m has not merged with other connected components, enter step d) to judge the intensity ratio; when the connected component P m has merged with other connected components and the number of merging points N ≤ 3, calculate the direction angles of all points; when N > 3, only use the last 3 merged points to calculate the direction angle. The direction angle formula is as follows:

[0117]

[0118] In the formula, (x l , y l ), (x l-1 , y l-1 ) are the coordinates of the merging points, 1 < l < L, where L is the length of the number of merged points. After that, add the point P to be incorporated m to (x1, y1)...(x L , y L ), and use the above formula to calculate the angle angle m after adding P Pre . When d angle = |angle Pre - angle Est | is less than the angle difference threshold A th , go to step d) for further judgment;​​​​​​​​​​​​​​​​​​​​​​​​​​​​​e) Loop on the next point and repeat steps b) to d) until all points have been traversed. Only points that meet the above conditions can be fused, and the center of the last fused point is used as the center of the fused point. Calculate the union of the fused points and use the union as the fused region;

[0123] f) Determine whether the connected domains after fusion intersect and further fuse the clusters.

[0124] 5. Underwater small target detection module

[0125] The idea is that the grayscale value of the target highlight is higher than the background area, and the grayscale value of the shadow is lower than the background area to detect small underwater targets. The specific process is as follows:

[0126] 1) Extract highlight, shadow and background areas. Remove areas larger than S from the connected domain. max and the area is less than S min The area of ​​​​the remaining connected domain is intercepted and the highlighted H (long h x , width h y ), shadow S area (length s x , width is h y ) and background B (length 2(h x +sx), width 2h y ). And calculate the average value B of the background area mean ;

[0127] 2) Calculate the pixel density of highlights and shadows. Let M H 、M S are the number of pixels in the highlight and shadow areas respectively; Ι is an indicator function, which is 1 when true and 0 otherwise. Calculate the highlight threshold τ h =ρ·B mean , according to the following formula, the density of highlighted pixels exceeding the threshold h c :

[0128]

[0129] At the same time, the shadow threshold is calculated as τ s =B mean / ρ, the pixel density of the shadow is as follows:

[0130]

[0131] 3) Using density weighted method, combined with highlights and shadows, calculate the detection score f of suspected targets c :

[0132] f c =ω c ·h c +(1-ω c )·sc

[0133] Where, ω c =0.7. When the detection score f c When it is greater than 0.5, it is a target to be identified.

[0134] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0135] Example 1

[0136] like Figure 1 As shown, the present invention discloses a method for detecting underwater small targets in sonar images under complex backgrounds, comprising the following steps: 1) denoising the sound image using multi-level median filtering; 2) extracting the highlight texture map and shadow texture map of the terrain through complex terrain masking; 3) significantly highlighting the highlight and shadow features using high-order moments, obtaining a threshold coefficient based on the false alarm rate, achieving highlight and shadow segmentation of the underwater small target, and then suppressing the background through the highlight texture map and shadow texture map; 4) performing association clustering on the processed areas using three types of detection windows: large, medium, and small, and effectively associating broken targets and suppressing false alarms through different clustering criteria; 5) using a statistical analysis method to combine the detection density of highlights and shadows, discriminate targets, and calculate the detection score of suspected small targets.

[0137] Figure 2 This is a schematic diagram of the acoustic image denoising process. A multi-level median filtering method is used to calculate the median of the azimuth and range directions of the image pixel, then calculate the median of the four corner points of the pixel, and finally replace the pixel with the median of these two values ​​and the pixel.

[0138] Figures 3(a) and 3(b) are schematic diagrams of background accumulation. Figure 3(a) is a schematic diagram of integral map calculation, that is, the value of any point on the map is the sum of the grayscale values ​​of all pixels in the rectangular area enclosed by the upper left corner of the image and the current point; Figure 3(b) is a background accumulation map, which contains the average of pixels in two rectangular areas. There may be a target in the middle area, so the middle area is removed.

[0139] Figure 4 This is a flow chart of high-order moment image segmentation. First, the acoustic image is exponentially transformed and the kurtosis value is calculated. Second, the highlight and shadow areas of the target are segmented according to the threshold. Third, the highlight and shadow images are filtered and smoothed respectively.

[0140] Figure 5 It is the highlight shadow position of the underwater small target acoustic image, where the red area is the highlight, the blue area is the shadow, and the green area is the background area. x Sum width = h y , the length of the shadow S area = s x Sum width = h yWhen the background B is centered on the highlight center, length = 2(h x +sx), width = 2h y .

[0141] Figure 6(a)-Figure 6(d) The results of large detection window processing are shown in Figure 6(a). Figure 6(b) shows the original sound image, Figure 6(c) shows the result after high-order moment segmentation and landform burial processing, Figure 6(d) shows the result after rule association. As can be seen from the figures, this method effectively suppresses false alarms caused by ridges.

[0142] Figure 7(a)-Figure 7(c) Figure 7(a) shows the target detection results of large-scale acoustic images. Figure 7(b) shows the detected connected domain of suspected targets. Figure 7(c) shows the score of target detection in large-scale acoustic images. When the detection score exceeds 0.5, the connected domain is a target.

[0143] Comparing Figures 7(a) and 7(c), the cable and float string are correctly detected. By combining highlight and shadow detection, two false alarms are suppressed.

[0144] Example 2

[0145] Embodiment 2 of the present invention proposes a system for detecting underwater small targets using sonar images under complex backgrounds. The system includes: a sonar image denoising processing module, a complex terrain mask processing module, a connected region clustering module, and a small target detection module. Based on the method of embodiment 1, the imaging sonar is used to collect acoustic images in real time, and acoustic image denoising, high-order moment image segmentation, and complex background mask processing are performed in sequence. After the two are correlated, a series of connected regions are obtained. Then, through connected region clustering processing, the broken targets are connected and the areas where the terrain residues are removed are removed. Finally, the underwater small targets are jointly identified based on the density ratio of the highlights and shadows of the original image corresponding to the remaining connected domains. Among them,

[0146] Sonar image denoising processing module, used to perform denoising on the collected sonar images using a multi-level median filtering method;

[0147] Complex terrain mask processing module, used to perform complex terrain mask processing on the denoised sonar image to extract the highlight texture map and shadow texture map of the terrain;

[0148] The high-order moment image segmentation module is used to perform high-order moment image segmentation on the denoised sonar image and perform background suppression by combining the results of the complex terrain mask processing module;

[0149] The connected region clustering module is used to cluster the connected regions processed by the high-order moment image segmentation module using three different sizes of detection windows, so as to achieve the association of split targets and the suppression of ridge landforms;

[0150] The small target detection module is used to extract highlight areas and shadow areas, calculate the highlight pixel density and shadow pixel density respectively, and obtain the detection score through pixel density weighted calculation to realize the detection of small targets.

[0151] Finally, it should be noted that the above embodiments are intended only to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the embodiments, it should be understood by those skilled in the art that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention and are intended to be encompassed by the claims of the present invention.

Claims

1. A method for detecting small underwater targets in sonar images under complex backgrounds, the method comprising: Step 1) adopting a multi-level median filtering method to denoise the collected sonar image; Step 2) performing complex terrain masking processing on the sonar image after denoising in step 1) to extract the highlight texture map and shadow texture map of the terrain; Step 3) performing high-order moment image segmentation on the sonar image after denoising in step 1), and combining the complex terrain mask processing result in step 2) to perform background suppression; Step 4) clustering the connected domains processed in step 3) using three different sizes of detection windows to achieve association of split targets and suppression of ridge landforms; Step 5) Extract highlight areas and shadow areas, calculate highlight pixel density and shadow pixel density respectively, and obtain detection scores through pixel density weighted calculation to achieve small target detection; The step 3) comprises: Step 3-1) performing a power exponential transformation on the sonar image de-noised in step 1); Step 3-2) Determine the neighborhood window Ω of any pixel (i, j) in the image G after power exponential transformation i,j , calculate the kurtosis value of the pixel HOG(i,j) according to the mean value in the neighborhood window: Where S i,j Represents the number of pixels in the neighborhood window, (p,q) is the neighborhood window Ω i,j Pixel coordinates within ; Step 3-3) Calculate the threshold coefficient ρ based on the false alarm rate, and achieve highlight and shadow segmentation by judging HOG(i,j): When the kurtosis value HOG(i,j) is less than ρ·η(p,q), the pixel belongs to the highlight area H T ; When the kurtosis value HOG(i,j) is greater than η(p,q) / ρ, the pixel belongs to the shadow area S T ; Step 3-4) Use morphological operators to remove the areas larger than S max and the area is smaller than S min area and perform edge smoothing; Step 3-5) Through the highlighted area H T , and the highlight texture image H obtained in step 2) B and shadow texture map S B Perform differential processing separately to achieve background suppression; The step 4) comprises: Step 4-1) Count the highlighted area graph H T Each connected domain in the network is found, and the area, center position and circumscribed rectangular frame of the connected domain are calculated. Step 4-2) Jointly count the pixel densities of highlights and shadows, and use a small detection window to determine whether adjacent connected domains are associated; Step 4-3) Using the middle detection window, determine whether the connected domains in the high-density area are associated; Step 4-4) By determining the strength and direction angle of the connected domain of the high-density area within the large detection window, the areas with distant distances including ridges and cables and scattered connected domains are associated.

2. The method for detecting underwater small targets in sonar images under complex background according to claim 1, characterized in that: The step 1) comprises: The sonar image is denoised by traversing each pixel of the image and calculating the median v of each pixel I(i,j) in the azimuth and range directions respectively. The pixel median u of the corner point where I(i,j) is located is then calculated. Finally, the median of I(i,j), v and u is used as the value of the pixel. Multi-level median processing is used to suppress noise and retain the edge contour of the target.

3. The method for detecting underwater small targets in sonar images under complex background according to claim 1, characterized in that: The step 2) comprises: Step 2-1) Using the integral graph method, background accumulation is performed on the sonar image after denoising in step 1) to obtain a cumulative graph B(i, j); Step 2-2) Using the statistical characteristics of the acoustic seabed topography that obeys the Rayleigh distribution, according to the required false alarm rate P fa Calculate the threshold V T : Where σ is the standard deviation of the Rayleigh distribution, and the threshold coefficient ρ is obtained from the following formula: Step 2-3) Using the cumulative image from step 2-1), perform large-scale sliding window processing and calculate the average value M(k) of the acoustic image in the k-th sliding window; when the cumulative image B(i,j)>M(k)·ρ, the pixel belongs to the highlight texture image H B ; When B(i,j)>M(k) / ρ, the pixel belongs to the shadow texture map S B ; Step 2-4) Use morphological operators to fill holes in the connected domains of the highlight texture map and the shadow texture map respectively; then remove the connected domains whose area is smaller than the preset area value S. max , the length or width is less than the preset length value L max The small target connected domain is obtained and the edges of the connected domain are smoothed.

4. The method for detecting underwater small targets in sonar images under complex background according to claim 1, characterized in that: The step 4-2) includes: Sort by the center position of the connected domains, then expand the upper, lower, left, and right pixels of the circumscribed rectangle of each connected domain. Determine whether it intersects with adjacent connected domains based on the border position relationship, and thus determine whether it is associated.

5. The method for detecting underwater small targets in sonar images under complex background according to claim 4, characterized in that: The step 4-3) includes: Step 4-3-1) Highlight the area H T Each connected region in is considered as a point P m , the position of the point is the center of the connected domain, m represents the sequence number of the point, and the statistics are based on point P m is the center and the side length is L med The number of points in the square area, when the number of points is greater than the set high-density area threshold, the area is a high-density area; Step 4-3-2) traverse each high-density area and determine whether adjacent high-density areas intersect. If the areas intersect, merge the intersecting areas to finally obtain a set of high-density areas; Step 4-3-3) From the highlighted area H T Remove the connected domains in high-density areas.

6. The method for detecting underwater small targets in sonar images under complex background according to claim 5, characterized in that: The step 4-4) includes: Step 4-4-1) Highlight the area H T The remaining connected domain is regarded as point P m , statistics are based on the side length L large The number of points in the square area is used to determine whether the area is a high-density area according to the methods of 4-3-1) and 4-3-2), and the high-density areas are merged and sorted according to the center coordinates of the merged high-density areas; Step 4-4-2) Calculate the vector connected domain P in the high-density area m and point P n When the distance is less than the threshold D th When , go to step 4-4-3); otherwise, loop to the next point; where n represents the serial number of the adjacent point; Step 4-4-3) When point P m Not fused with other points, go to step 4-4-4) to judge the intensity ratio; when point P m It has been fused with other points, and the number of fused points is less than 3. Calculate the direction angle of all point differences. Est ; When the number of fusion points is greater than or equal to 3, only the last fused 3 points are used to calculate the direction angle angle Est , then calculate the point P to be integrated n Angle after adding Pre , when d angle =|angle Pre -angle Est |Less than the angle difference threshold A th If yes, go to step 4-4-4) for further judgment; otherwise, loop to the next point; Step 4-4-4) Calculate the two adjacent connected regions P m and P n The corresponding average intensity in the original image, and calculate its intensity ratio R meanI (m,n): When the intensity ratio R meanI (m,n) is greater than the threshold R th When P m and P n Fusion; among them, meanI m and meanI n Represents point P respectively m and P n The connected domain referred to corresponds to the average intensity value in the original image; Step 4-4-5) loops for the next point and goes to step 4-4-2) until all points are traversed and goes to step 4-4-6); Step 4-4-6) Take the center of the last fused point as the center of the fused point, calculate the union of the fused points, and use the union as the fused area; Step 4-4-7) Determine whether each connected domain after fusion intersects and further fuse the clusters.

7. The method for detecting underwater small targets in sonar images under complex background according to claim 5, characterized in that: The step 5) comprises: Step 5-1) Remove the areas that are too large and too small from the connected domain, and intercept the highlight area, shadow area and background area in the original image where the remaining connected domain is located, and calculate the average value B of the background area mean ; Step 5-2) Calculate the pixel density of the highlight area and the shadow area; Step 5-3) Based on the above pixel density, a weighted average method is used to calculate the detection score f of the connected domain. c ; Compare with the set value to complete small target detection.

8. A system for detecting small underwater targets in sonar images under complex backgrounds, implemented according to the method described in any one of claims 1 to 7, characterized in that: The system includes: a sonar image denoising processing module, a complex terrain mask processing module, a high-order moment image segmentation module, a connected region clustering module and a small target detection module; wherein, The sonar image denoising processing module is used to perform denoising on the collected sonar images using a multi-stage median filtering method; The complex terrain mask processing module is used to perform complex terrain mask processing on the denoised sonar image to extract the highlight texture map and shadow texture map of the terrain; The high-order moment image segmentation module is used to perform high-order moment image segmentation on the denoised sonar image and perform background suppression in combination with the results of the complex terrain mask processing module; The connected region clustering module is used to cluster the connected regions processed by the high-order moment image segmentation module using three different-sized detection windows, so as to achieve the association of split targets and the suppression of ridge landforms; The small target detection module is used to extract highlight areas and shadow areas, calculate highlight pixel density and shadow pixel density respectively, and obtain a detection score through pixel density weighted calculation to achieve small target detection.

Images