Fusion detection method for underwater small target

Abstract

The invention provides a fusion detection method for an underwater small target, and the method comprises the steps: 1, obtaining a sonar image of an underwater target, carrying out the preprocessing of the sonar image, and obtaining a preprocessed image; step 2, target detection is carried out on the preprocessed image by adopting an FCM detection method and a k-means detection method respectively, preliminary detection results of the FCM detection method and the k-means detection method are obtained respectively, and the preliminary detection results are binary or ternary detection images; step 3, clustering the preliminary detection results by using a morphological spatial clustering method, and respectively obtaining final detection results of an FCM detection method and a k-means detection method; and step 4, carrying out detection result fusion on the final detection result by adopting a fusion strategy, and completing fusion detection for the underwater small target. The underwater weak and small target automatic detection of the high-resolution imaging sonar is realized under the condition of lower signal-to-noise ratio, and the method has the characteristic of remarkable low false alarm rate.

Description

Technical Field

[0001] The invention relates to a target detection method, in particular to a fusion detection method for underwater small targets. Background Art

[0002] This section merely provides background information related to the present disclosure and is not necessarily prior art.

[0003] Underwater target detection needs to solve the problem of existence and non-existence. The detection algorithm detects underwater targets based on the sound waves reflected from the sonar image. At present, it is generally divided into the following three types: pixel-based underwater target detection, feature-based underwater target detection, and echo-based underwater target detection. Common pixel-based underwater target detection algorithms include frame difference method, optical flow method, and background subtraction method; feature-based underwater target detection is a detection algorithm that detects based on the significant features of the target, such as symbol analysis method; echo-based underwater target detection uses the physical properties of different objects to cause different echoes, and thus realizes the detection of buried targets on the seabed. Traditional underwater target detection algorithms generally have the problem of high false alarm rate or high missed detection rate. Therefore, reducing the false alarm rate while still maintaining a high probability of correct detection is one of the problems that need to be solved in underwater target detection.

[0004] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present disclosure, and therefore may include information that does not constitute the prior art known to ordinary technicians in the field. Summary of the invention

[0005] Purpose of the invention: The technical problem to be solved by the present invention is to provide a fusion detection method for small underwater targets in view of the shortcomings of the prior art.

[0006] In order to solve the above technical problems, the present invention discloses a fusion detection method for underwater small targets, comprising the following steps:

[0007] Step 1, obtaining a sonar image of an underwater target, and preprocessing it to obtain a preprocessed image;

[0008] Step 2, using the preprocessed image to perform target detection using the FCM detection method and the k-means detection method, respectively, to obtain preliminary detection results of the FCM detection method and the k-means detection method, respectively, wherein the preliminary detection results are binary or ternary detection images;

[0009] Step 3, clustering the preliminary detection results using a morphological space clustering method to obtain final detection results of the FCM detection method and the k-means detection method respectively;

[0010] Step 4: adopt a fusion strategy to fuse the final detection results to obtain a fused detection result, thereby completing the fusion detection for underwater small targets.

[0011] Further, the pre-processing described in step 1 specifically includes:

[0012] Step 1-1, downsampling the sonar image in the distance direction using a one-dimensional linear interpolation method;

[0013] Step 1-2, using Gamma grayscale correction to suppress the background noise of the sonar image.

[0014] Furthermore, downsampling the sonar image in step 1-1 specifically includes:

[0015] Step 1-1-1, determining a downsampling multiple according to the size of the sonar image, and reducing the number of pixels in the distance direction to a preset multiple of the beam direction;

[0016] Step 1-1-2, according to the downsampling multiple, one-dimensional linear interpolation downsampling is performed in the distance direction, and the horizontal and vertical coordinates are converted to the actual beam angle and radial distance. The process of linear interpolation downsampling is expressed as follows:

[0017]

[0018] Among them, x 0 , x 1 Indicates the distance between two adjacent sampling points under a fixed beam; y 0 ,y 1 Indicates that at distance x 0 , x 1 The upsampled pixel value; x represents the downsampled distance value, and y represents the pixel value calculated at distance x.

[0019] Furthermore, the Gamma grayscale correction described in step 1-2 specifically includes:

[0020] Step 1-2-1, set parameters, set the value of Gamma grayscale correction parameter γ;

[0021] Step 1-2-2, performing a Gamma transformation on the sonar image downsampled in step 1-1 with a set γ value to obtain a preprocessed image; wherein the Gamma transformation is expressed as follows:

[0022] s=(r+c) γ r∈[0,1]

[0023] Among them, r is the normalized original grayscale value; s is the transformed grayscale value; c is the compensation coefficient; the parameter γ is used to describe the curve shape of the relationship between the input and output values, and is a preset non-negative scalar.

[0024] Furthermore, the target detection described in step 2 specifically includes:

[0025] Step 2-1, using the FCM detection method to perform target detection on the preprocessed image, specifically includes the following steps:

[0026] Step 2-1-1, converting the preprocessed image into a membership graph, as follows:

[0027] A first cost function is constructed, and the conversion is performed with the goal of minimizing the first cost function, wherein the first cost function J1 is expressed as follows:

[0028]

[0029] Among them, i n is the gray value of the nth pixel in the preprocessed image; N is the total number of pixels; C is the number of clusters; v m is the centroid of the mth class; ‖‖ represents the norm; represents the membership function of the nth pixel to the mth class; l>1.1 is the fuzzy weighted index;

[0030] Membership function μ mn Subject to the following constraints:

[0031]

[0032] 0≤μ mn ≤1

[0033]

[0034] Using the above method, the membership degree of any pixel in the image in any cluster formed after clustering is normalized to a range of 0 to 1;

[0035] Using the Lagrangian method, the membership function μ mn Minimize the first cost function J1 under the constraint of , and get the membership function μ mn and the centroid v m The update formula is as follows:

[0036]

[0037]

[0038] Thus, C membership graphs having the same size as the sonar image are obtained;

[0039] Step 2-1-2, set the threshold and refine the grouping to obtain a binary or ternary detection image as the preliminary detection result of the FCM detection method, as follows:

[0040] Step 2-1-2-1, set the threshold, pre-set the membership threshold d0, d0∈(0,1);

[0041] Step 2-1-2-2, refine the grouping, adopt the refinement grouping strategy, that is, set the number of clusters C greater than or equal to 2 or 3, all pixels are divided into the 1st, ..., C categories according to the gray value from low to high, the 1st category is set as the shadow area, the Cth category is set as the bright area, and the other categories are backgrounds, and a binary or ternary detection image is obtained;

[0042] Step 2-2, using the k-means detection method to perform target detection on the preprocessed image, specifically includes the following steps:

[0043] Step 2-2-1, construct the second cost function J2, as follows:

[0044]

[0045] Target detection is performed by minimizing the second cost function J2;

[0046] Step 2-2-2, refine the grouping and obtain the binary or ternary detection image as the preliminary detection result of the k-means detection method, as follows:

[0047] Use the method described in step 2-2-1 to perform target detection, that is, divide the image into K clusters, where K>NUN, NUN=2 or 3, indicating a binary or ternary image, and label the K clusters from low to high according to the grayscale value:

[0048] label=1,2,…,K

[0049] Set a detection threshold d, where d∈label and the length of d is NUM-1. Use the detection threshold d to divide the K clusters into NUM segments to obtain a binary or ternary detection image.

[0050] Furthermore, the morphological spatial clustering method described in step 3 specifically includes the following steps:

[0051] Step 3-1: For the input image f(x,y), the structural element k(i,j) is used to perform the dilation operation, as follows:

[0052]

[0053] in, represents the dilation operation, (i, j) represents the coordinates of the structural element k, and (x, y) represents the pixel coordinates on the image f;

[0054] Step 3-2, through the expansion operation described in step 3-1, the regions of interest ROIs of the preliminary detection results of the FCM detection method and the k-means detection method are merged, and the ROIs are screened and spatially clustered according to prior knowledge, and finally three types of clusters are formed, as follows:

[0055] The first type, individual highlight clusters, consist of one or more associated highlighted regions of interest ROIs;

[0056] The second type, individual shadow clusters, consist of one or more associated shadow regions of interest ROIs;

[0057] The third type, associated highlight and shadow clusters, consists of one or more highlight regions of interest ROIs and one or more shadow regions of interest ROIs.

[0058] Furthermore, the screening and spatial clustering according to prior knowledge described in step 3-2 specifically include:

[0059] An all-1 matrix is ​​selected as a structural element, a preset ratio of the sonar image distance is selected as a size, and a preset ratio of the beam angle is selected for screening and spatial clustering.

[0060] Furthermore, the detection result fusion described in step 4 specifically includes:

[0061] Step 4-1, positionally associate the final detection results of the FCM detection method and the k-means detection method obtained in step 3, i.e., ROIs;

[0062] In step 4-2, the two final detection results, i.e., the ROIs whose areas overlap by more than a preset percentage, are taken as fusion detection results.

[0063] Furthermore, the preset multiple described in step 1-1-1 is 10 times;

[0064] The value of the Gamma grayscale correction parameter γ described in step 1-2-1 is 1.2;

[0065] The membership threshold d0 described in step 2-1-2-1 has a default value of 0.5.

[0066] Further, the preset ratio of the sonar image distance described in step 3-2 is 3%;

[0067] The preset ratio of the beam angle described in step 3-2 does not exceed 2.5%;

[0068] The preset ratio of the beam angle described in step 3-2 is 2%;

[0069] The preset percentage described in step 4-2 is 50%.

[0070] Beneficial effects:

[0071] 1. The present invention overcomes the sensitivity of the performance of a single target detection method to parameters related to the marine environment and to the signal processing method used to generate the acoustic image.

[0072] 2. In the fusion detection method proposed in the present invention, multiple independent methods process the same data and then fuse their outputs into a single output. The fusion output is usually more robust than the output of any single detection method, reducing the false alarm rate while maintaining a high probability of correct detection.

[0073] 3. The fusion processing method proposed in the present invention eliminates the irrelevance of false alarms between different methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments, and the above and / or other advantages of the present invention will become more clear.

[0075] Figure 1 It is a schematic diagram of the overall process of the method of the present invention.

[0076] Figure 2a This is a schematic diagram of the results of FCM detection + morphological spatial clustering.

[0077] Figure 2b This is a schematic diagram of the final FCM detection results.

[0078] Figure 3a This is a schematic diagram of the FCM and k-means fusion detection results in actual situations.

[0079] Figure 3b This is another schematic diagram of FCM and k-means fusion detection results in actual situations. DETAILED DESCRIPTION

[0080] The present invention proposes a fusion detection method for underwater small targets, comprising the following steps:

[0081] Step 1: Downsampling. Considering that the increase in the amount of sonar image data is usually caused by the increase in the number of sampling points, and the number of beams is relatively small and fixed, one-dimensional linear interpolation is used to downsample in the distance direction. This can shorten the processing time and avoid reducing the resolution in the beam direction. The interpolation process can be expressed as:

[0082]

[0083] Among them, g s (lΔx, kΔy) is the discrete expression of the original image; h n () is the interpolation kernel function.

[0084] Step 2: Grayscale correction. Generally, the grayscale value distribution of sonar images is relatively concentrated, that is, the grayscale variation range is small. From the grayscale histogram, a small part of the grayscale value occupies the vast majority of the number of pixels. Through Gamma grayscale correction, the target can be effectively highlighted in the background. The basic form of Gamma transformation is:

[0085] s=(r+c) γ r∈[0,1]

[0086] Among them, r is the normalized original grayscale value; s is the transformed grayscale value; c is the compensation coefficient; γ describes the curve shape of the relationship between the input and output values ​​and is a non-negative scalar.

[0087] Step 3: Target detection: The preprocessed sonar images are subjected to FCM and k-means detection to obtain binary / ternary detection images. The FCM detection results and k-means detection results are obtained through morphological space clustering. Finally, the detection is fused under the 2 / 2 strategy to output the fused detection results.

[0088] Further, in step 1, the downsampling comprises the following steps:

[0089] Step 1-1, determine the downsampling multiple according to the original image size, that is, the number of sampling points × the number of beams;

[0090] Step 1-2: perform one-dimensional linear interpolation downsampling in the distance direction by the multiple set in step 1-1, and convert the horizontal and vertical coordinates to the actual beam angle and radial distance.

[0091] Furthermore, in step 2, the Gamma grayscale correction specifically includes the following steps:

[0092] Step 2-1, select a suitable γ value, generally around 1.2. A value that is too large or too small may cause image distortion.

[0093] Step 2-2: Perform a gamma transformation on the downsampled sonar image using the gamma value selected in step 2-1 to suppress background noise and highlight the target.

[0094] Furthermore, in step 3, the multi-algorithm fusion target detection specifically includes the following steps:

[0095] Step 3-1, k-means clustering detection, the algorithm principle is to try to minimize the cost function:

[0096]

[0097] Among them, i n is a specific image pixel; v m is the centroid of the mth cluster; ‖‖ represents the norm. The ideal result of the k-means algorithm is to maximize the variation between classes and minimize the variation within clusters.

[0098] Like many other types of numerical minimization, the solution obtained by k-means depends largely on the starting point, that is, the k-means algorithm converges to a local (non-global) minimum solution, which may exist in other better solutions. To improve this problem, a strategy of refining the grouping and re-merging is adopted. For example, if n=3 groups of bright area, shadow area and background (H / S / B) are desired, the image is first divided into K (K>n) clusters using k-means, and labels are labeled label=1,2,…,K from low to high grayscale values, and then the K clusters are divided into n segments (groups) by a pre-set detection threshold d (d∈label and the length of d is n-1);

[0099] Step 3-2: Fuzzy c-means (FCM) clustering detection, the algorithm principle is to minimize the cost function:

[0100]

[0101] Among them, i n is the gray value of the nth pixel in the image; N is the total number of pixels; C is the number of clusters; v m is the centroid of the mth class; μ mn It represents the membership of the n-th pixel to the m-th category; l (>1.1) is the fuzzy weighted index, and the larger the value of l, the higher the degree of fuzziness.

[0102] Membership function μ mn Subject to the following constraints:

[0103]

[0104] 0≤μ mn ≤1

[0105]

[0106] Normalize the membership of any pixel in any cluster to 0-1.

[0107] Using the Lagrangian method in μ mn By minimizing J under the constraints ofmn and the centroid v m The update formula is:

[0108]

[0109]

[0110] Obviously, pixels close to the centroid are assigned high membership values, while pixels far from the centroid are assigned low membership values.

[0111] In order to minimize the cost function, the FCM algorithm adaptively estimates the centroid and scope of each subclass, so it can be used as an adaptive threshold detection segmentation method. Since the direct result of FCM is C membership graphs of the same size as the original image, where C is the number of clusters, the larger the value of a pixel in the image in the membership graph of a cluster, the higher the probability that the pixel belongs to this cluster, so a membership threshold d0 (d0∈(0,1), the default value is 0.5) should be set in advance. For a category of interest, the points in its membership graph that are greater than d0 are used as detection outputs.

[0112] In the process of sonar image processing, the H / B binary grouping of bright area and background, or the H / S / B ternary grouping of bright area, shadow and background is generally used. Considering that the dynamic range of the background in the sonar image is much larger than that of the target, or the contrast between the target and the background is not high enough, a refined grouping strategy is adopted, that is, C is greater than or equal to 2 or 3, and the pixel clusters are divided into 1, 2...C categories from low to high grayscale values. The 1st category is the shadow area, the Cth category is the bright area, and the rest are backgrounds.

[0113] Step 3-3, morphological processing, for the image f(x, y), the formula for the expansion operation using the structural element k(x, y) is:

[0114]

[0115] The results directly detected by k-means or FCM often contain a lot of interference information, such as a target being split into several, large echoes or small noise points. The ROIs detected by k-means or FCM are merged through morphological expansion operations, and then some ROIs with particularly large or small areas are screened out based on prior knowledge such as target scale, so that the number of interference is significantly reduced while the target area is retained. For simplicity of operation, the structural element selects a full 1 matrix, and the size is selected as 3% of the original sonar image distance and 2% of the beam angle (due to the relatively small number of beams, it is best not to exceed 2.5% to avoid excessive merging). The ROIs that have been spatially clustered eventually form the following three types of clusters.

[0116] Type 1: Individual highlight clusters, consisting of one or more associated highlight ROIs.

[0117] Type 2: Individual shadow clusters, consisting of one or more associated shadow ROIs.

[0118] Type 3: Associated highlight and shadow clusters, which consist of one or more highlight ROIs and one or more shadow ROIs.

[0119] Step 3-4: Detection algorithm fusion processing, the detection results of FCM and k-means after morphological processing, that is, ROIs, are associated with each other. The ROIs detected by both algorithms are considered potential targets, and other ROIs are considered interference. This is actually a 2-choose-2 voting process. In actual operation, ROIs with an area overlap of more than 50% are considered to be the same target.

[0120] Example:

[0121] like Figure 1 As shown, the specific implementation of the underwater small target detection method based on multi-method fusion involved in the present invention mainly includes two parts: preprocessing and multi-algorithm fusion detection. The preprocessing mainly includes two steps of downsampling and Gamma grayscale correction, and the multi-algorithm fusion detection mainly includes three steps of FCM+morphological processing clustering detection, k-means+morphological processing clustering detection, and 2 / 2 fusion detection.

[0122] First, the downsampling multiple is determined according to the original image size, that is, the number of sampling points × the number of beams. One-dimensional linear interpolation downsampling is performed in the distance direction, and the horizontal and vertical coordinates are converted to the actual beam angle and radial distance.

[0123] Then, select an appropriate γ value based on experience, generally around 1.2. A value that is too large or too small can easily cause image distortion. Perform a gamma transform on the downsampled sonar image to suppress background noise, highlight the target, and complete the sonar image preprocessing.

[0124] Then, FCM+morphological processing cluster detection and k-means+morphological processing cluster detection are performed in parallel. The specific implementation process of k-means cluster detection is as follows:

[0125] Input: sonar image I, segmentation number N (=2 or 3), clustering number K (≥N), detection threshold d.

[0126] Output: k-means detection result map mask.

[0127] 1: Calculate the grayscale histogram of sonar image I.

[0128] 2: Initialize the centroid: take K samples at equal intervals from 0 to the maximum grayscale value (following the principle that the initial centroid distance is as far as possible).

[0129] 3: Set the iteration termination conditions: the maximum number of iterations and the distance threshold Δ between the two cluster centers.

[0130] 4: while (the distance between two adjacent centroids > Δ| reaches the maximum number of iterations), do

[0131] 5: Calculate the cluster label of the grayscale value according to formula (3-1): the grayscale value belongs to the cluster that is closest to the centroid.

[0132] 6: Update the centroid: weighted average the grayscale values ​​within the cluster according to the grayscale histogram.

[0133] 7: Arrange the centroids from small to large and calculate the distance between the two centroids.

[0134] 8: end while

[0135] 9: if N = 2 (detect bright area, d has 1 value)

[0136] 10: mask: Set pixels whose label is greater than or equal to d to 1 (bright area), and set the rest of the pixels to 0 (background).

[0137] 11: else (N=3, detect bright area and shadow area, d has 2 values)

[0138] 12: mask: pixels with label greater than or equal to d(1) are set to 1 (bright area), pixels with label less than or equal to d(2) are set to -1 (shadow area), and the rest of the pixels are set to 0 (background).

[0139] 13: end if

[0140] The specific implementation process of FCM cluster detection is as follows:

[0141] Input: sonar image I, number of segmentations N (=2 or 3), number of clusters C (≥N).

[0142] Output: FCM detection result map mask.

[0143] 1: Set the parameter default values: membership threshold d0=0.5, fuzzy weighting index l=2.

[0144] 2: Calculate the grayscale histogram of sonar image I.

[0145] 3: Initialize the centroid: take K samples at equal intervals from 0 to the maximum grayscale value (following the principle that the initial centroid distance is as far as possible).

[0146] 4: Set the iteration termination condition: the distance threshold Δ between the two cluster centers.

[0147] 5: while (the distance between the two adjacent centroids > Δ), do

[0148] 6: Calculate the fuzzy membership according to formula (3-4).

[0149] 7: Update the center of mass according to formula (3-5).

[0150] 8: Arrange the centroids from small to large and calculate the distance between the two centroids.

[0151] 9: end while

[0152] 10: if N = 2 (detect bright area)

[0153] 11: mask: Extract the membership map of class C, and set the pixels greater than or equal to d0 to 1 (bright area) and the rest to 0 (background).

[0154] 12: else (N=3, detect bright area and shadow area)

[0155] 13: mask: extract the membership map of class C, and set the pixels greater than or equal to d0 to 1 (bright area); extract the membership map of class 1, and set the pixels greater than or equal to d0 to -1 (shadow area); the remaining pixels are set to 0 (background).

[0156] 14: end if

[0157] Finally, a 2 / 2 fusion detection of FCM and k-means is performed. The specific operation is to positionally associate the detection results of FCM and k-means after morphological processing, namely ROIs. It is considered that the ROIs detected by both algorithms (the ROI area overlap exceeds 50%) are potential targets, and other ROIs are considered to be interference.

[0158] Figure 2a and Figure 2b The detection process of a single detection algorithm FCM is shown, where Figure 2a It is the result of FCM detection + morphological spatial clustering, which uses the property that the detection points belonging to the same object are continuous in space to reduce the number of detection points. Figure 2b The final FCM detection result is obtained by area screening and deleting ROIs that are too large or too small.

[0159] Figure 3a and Figure 3bThe results of FCM and k-means fusion detection are shown, where the rectangles are ROIs detected by FCM, the circles are ROIs detected by k-means, and the triangles are ROIs after FCM and k-means fusion processing. To verify the effectiveness of the algorithm, here Figure 3a and Figure 3b Two groups of sonar image data with different targets, different sonars and different waters were selected. Figure 3a There are two real cylindrical targets at (5.5°, 180m) and (7°, 140m). Figure 3b There is a real cube frame target at (1.1°, 23.2m). Obviously, the fusion processing effectively reduces the interference without missing the real target.

[0160] In a specific implementation, the present application provides a computer storage medium and a corresponding data processing unit, wherein the computer storage medium can store a computer program, and when the computer program is executed by the data processing unit, the invention content of a fusion detection method for underwater small targets provided by the present invention and some or all steps in each embodiment can be executed. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0161] Those skilled in the art can clearly understand that the technical solutions in the embodiments of the present invention can be implemented by means of computer programs and their corresponding general hardware platforms. Based on such an understanding, the technical solutions in the embodiments of the present invention can be essentially or partly contributed to the prior art in the form of computer programs, i.e., software products, which can be stored in a storage medium and include several instructions for enabling a device including a data processing unit (which can be a personal computer, a server, a single-chip microcomputer, an MCU or a network device, etc.) to execute the methods described in various embodiments of the present invention or certain parts of the embodiments.

[0162] The present invention provides a method and idea for a fusion detection method for underwater small targets. There are many methods and ways to implement the technical solution. The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, without departing from the principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention. All components not specified in this embodiment can be implemented by existing technologies.

Claims

1. A fusion detection method for underwater small targets, characterized in that: The following steps are involved: Step 1, obtaining a sonar image of an underwater target, and preprocessing it to obtain a preprocessed image; Step 2, using the preprocessed image to perform target detection using the FCM detection method and the k-means detection method, respectively, to obtain preliminary detection results of the FCM detection method and the k-means detection method, respectively, wherein the preliminary detection results are binary or ternary detection images; Step 3, clustering the preliminary detection results using a morphological space clustering method to obtain final detection results of the FCM detection method and the k-means detection method respectively; Step 4: adopt a fusion strategy to fuse the final detection results to obtain a fused detection result, thereby completing the fusion detection for underwater small targets.

2. A fusion detection method for underwater small targets according to claim 1, characterized in that: The pre-processing described in step 1 specifically includes: Step 1-1, downsampling the sonar image in the distance direction using a one-dimensional linear interpolation method; Step 1-2, using Gamma grayscale correction to suppress the background noise of the sonar image.

3. A fusion detection method for underwater small targets according to claim 2, characterized in that: The downsampling of the sonar image in step 1-1 specifically includes: Step 1-1-1, determining a downsampling multiple according to the size of the sonar image, and reducing the number of pixels in the distance direction to a preset multiple of the beam direction; Step 1-1-2, according to the downsampling multiple, one-dimensional linear interpolation downsampling is performed in the distance direction, and the horizontal and vertical coordinates are converted to the actual beam angle and radial distance. The process of linear interpolation downsampling is expressed as follows: Among them, x0 and x1 represent the distance values ​​of two adjacent sampling points under the fixed beam; y0 and y1 represent the pixel values ​​sampled at the distances x0 and x1; x represents the distance value after downsampling, and y represents the pixel value calculated at the distance x.

4. A fusion detection method for underwater small targets according to claim 3, characterized in that: The gamma grayscale correction described in step 1-2 specifically includes: Step 1-2-1, set parameters, set the value of Gamma grayscale correction parameter γ; Step 1-2-2, performing a Gamma transformation on the sonar image downsampled in step 1-1 with a set γ value to obtain a preprocessed image; wherein the Gamma transformation is expressed as follows: s=(r+c) γ r∈[0,1] Among them, r is the normalized original grayscale value; s is the transformed grayscale value; c is the compensation coefficient; the parameter γ is used to describe the curve shape of the relationship between the input and output values, and is a preset non-negative scalar.

5. A fusion detection method for underwater small targets according to claim 4, characterized in that: The target detection described in step 2 specifically includes: Step 2-1, using the FCM detection method to perform target detection on the preprocessed image, specifically includes the following steps: Step 2-1-1, converting the preprocessed image into a membership graph, as follows: A first cost function is constructed, and the conversion is performed with the goal of minimizing the first cost function, wherein the first cost function J1 is expressed as follows: Among them, i n is the gray value of the nth pixel in the preprocessed image; N is the total number of pixels; C is the number of clusters; v m is the centroid of the mth class; ‖‖ represents the norm; represents the membership function of the nth pixel to the mth class; l>1.1 is the fuzzy weighted index; Membership function μ mn Subject to the following constraints: 0≤μ mn ≤1 Using the above method, the membership degree of any pixel in the image in any cluster formed after clustering is normalized to a range of 0 to 1; Using the Lagrangian method, the membership function μ mn Minimize the first cost function J1 under the constraint of , and get the membership function μ mn and the centroid v m The update formula is as follows: Thus, C membership graphs having the same size as the sonar image are obtained; Step 2-1-2, set the threshold and refine the grouping to obtain a binary or ternary detection image as the preliminary detection result of the FCM detection method, as follows: Step 2-1-2-1, set the threshold, pre-set the membership threshold d0, d0∈(0,1); Step 2-1-2-2, refine the grouping, adopt the refinement grouping strategy, that is, set the number of clusters C greater than or equal to 2 or 3, all pixels are divided into the 1st, ..., C categories according to the gray value from low to high, the 1st category is set as the shadow area, the Cth category is set as the bright area, and the other categories are backgrounds, and a binary or ternary detection image is obtained; Step 2-2, using the k-means detection method to perform target detection on the preprocessed image, specifically includes the following steps: Step 2-2-1, construct the second cost function J2, as follows: Target detection is performed by minimizing the second cost function J2; Step 2-2-2, refine the grouping and obtain the binary or ternary detection image as the preliminary detection result of the k-means detection method, as follows: Use the method described in step 2-2-1 to perform target detection, that is, divide the image into K clusters, where K>NUM, NUM=2 or 3, indicating a binary or ternary image, and label the K clusters from low to high according to the grayscale value: label=1,2,…,K Set a detection threshold d, where d∈label and the length of d is NUM-1. Use the detection threshold d to divide the K clusters into NUM segments to obtain a binary or ternary detection image.

6. A fusion detection method for underwater small targets according to claim 5, characterized in that: The morphological spatial clustering method described in step 3 specifically includes the following steps: Step 3-1: For the input image f(x,y), the structural element k(i,j) is used to perform the dilation operation, as follows: in, represents the dilation operation, (i, j) represents the coordinates of the structural element k, and (x, y) represents the pixel coordinates on the image f; Step 3-2, through the expansion operation described in step 3-1, the regions of interest ROIs of the preliminary detection results of the FCM detection method and the k-means detection method are merged, and the ROIs are screened and spatially clustered according to prior knowledge, and finally three types of clusters are formed, as follows: The first type, individual highlight clusters, consist of one or more associated highlighted regions of interest ROIs; The second type, individual shadow clusters, consist of one or more associated shadow regions of interest ROIs; The third type, associated highlight and shadow clusters, consists of one or more highlight regions of interest ROIs and one or more shadow regions of interest ROIs.

7. A fusion detection method for underwater small targets according to claim 6, characterized in that: The screening and spatial clustering based on prior knowledge described in step 3-2 specifically include: An all-1 matrix is ​​selected as a structural element, a preset ratio of the sonar image distance is selected as a size, and a preset ratio of the beam angle is selected for screening and spatial clustering.

8. A fusion detection method for underwater small targets according to claim 7, characterized in that: The detection result fusion described in step 4 specifically includes: Step 4-1, positionally associate the final detection results of the FCM detection method and the k-means detection method obtained in step 3, i.e., ROIs; In step 4-2, the two final detection results, i.e., the ROIs whose areas overlap by more than a preset percentage, are taken as fusion detection results.

9. A fusion detection method for underwater small targets according to claim 8, characterized in that: The preset multiple described in step 1-1-1 is 10 times; The value of the gamma grayscale correction parameter γ described in step 1-2-1 is 1.2; The membership threshold d0 described in step 2-1-2-1 has a default value of 0.

5.

10. The fusion detection method for underwater small targets according to claim 8, characterized in that: The preset ratio of the sonar image distance described in step 3-2 is 3%; The preset ratio of the beam angle described in step 3-2 does not exceed 2.5%; The preset ratio of the beam angle described in step 3-2 is 2%; The preset percentage described in step 4-2 is 50%.

Images