Forward-looking sonar image small target detection and recognition algorithm

Abstract

The invention belongs to the technical field of sonar image small target detection, and particularly relates to a foresight sonar image small target detection and recognition algorithm. Comprising the following steps: step 1, inputting sonar beam domain data; 2, detecting the position of a small target through a constant false alarm algorithm; step 3, extracting target HOG features of the sonar image through an HOG algorithm; and 4, identifying the type of the detected target through an SVM algorithm, and carrying out image classification. The algorithm is high in operation efficiency, good in timeliness and good in detection and recognition capability.

Description

Technical Field

[0001] The invention belongs to the technical field of sonar image small target detection, and in particular relates to a forward-looking sonar image small target detection and recognition algorithm. Background Art

[0002] Forward-looking sonar is an important instrument for ocean target detection, especially small target detection, and is mainly used for accurate positioning and classification of close-range targets in the ocean. For sonar target monitoring, it is difficult for sonar monitoring personnel to conduct long-term and efficient monitoring. With the help of computer technology, signal processing technology, artificial intelligence technology, etc., maximizing the automatic detection, tracking, identification and alarm of targets as much as possible is the development direction of sonar target detection, tracking and monitoring. Therefore, the small target detection and recognition algorithm can be used to achieve the purpose of reducing false alarms and improving autonomous recognition capabilities, providing users with direct and clear small target detection and recognition information. Summary of the invention

[0003] The purpose of the present invention is to overcome the above-mentioned shortcomings of the prior art. The present invention provides a small target detection and recognition algorithm for forward-looking sonar images, which is used to automatically detect small targets in forward-looking sonar images and automatically identify target types. The target detection and recognition information is uploaded to the user end to provide the user with small target detection and recognition results.

[0004] To solve the above technical problems, the present invention provides a small target detection and recognition algorithm for forward-looking sonar images, comprising the following steps:

[0005] Step 1: Sonar beam domain data input;

[0006] Step 2: Detect the position of small targets through the constant false alarm algorithm;

[0007] Step 3: Extract the target HOG features of the sonar image through the HOG algorithm;

[0008] Step 4: Use the SVM algorithm to identify the type of detected targets and perform image classification.

[0009] Preferably, in step 1, the sonar beam domain data is input to a constant false alarm algorithm for detection, and the input is raw data output by the beamforming, and the output of the constant false alarm algorithm for detection is target detection binary data.

[0010] Preferably, in step 2, the constant false alarm algorithm detection is used to detect the target echo part in the echo signal, and the constant false alarm algorithm detection consists of two major functional parts: background estimation and target detection; that is, after data input, background estimation is performed first, and then target detection is performed.

[0011] Preferably, before extracting the target HOG features in step 3, it also includes: image preprocessing, that is: first, when used with a classifier, the image training set needs to be resized, the size of which depends on the target size and the sonar image resolution, and the target only occupies the main part of the sample image; secondly, the sample image is gamma corrected.

[0012] Preferably, the target HOG feature extraction in step 3 specifically includes the following steps:

[0013] The first step is to divide the sample image into several pixel units, and then divide the gradient direction into 9 intervals. In each unit, the gradient directions of all pixels in each direction interval are histogram-counted to obtain a 9-dimensional feature vector.

[0014] In the second step, every 4 adjacent units form a block, and the feature vectors in a block are concatenated to obtain a 36-dimensional feature vector and normalized;

[0015] The third step is to scan the sample image with blocks, with a scanning step size of one unit. Finally, the features of all blocks are connected in series to obtain a feature vector of a sample.

[0016] Preferably, the image classification in step 4 specifically includes the following steps:

[0017] First, determine the SVM kernel function as a linear kernel and configure the SVM algorithm parameters;

[0018] Secondly, the optimal condition coefficients are iterated for the binary labeled HOG feature vector group and the decision plane intercept B is calculated to obtain the SVM model;

[0019] Finally, other binary labeled HOG feature vector groups are used to test and verify the obtained SVM model.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] The present invention proposes a small target detection and recognition algorithm for forward-looking sonar images based on the constant false alarm algorithm and the HOG-SVM algorithm. The algorithm first detects the position of the small target by combining the sonar beam domain data with the constant false alarm algorithm, then identifies the target type of the detected target by the HOG-SVM algorithm, and finally integrates the detection and recognition information. The algorithm has high computational efficiency, good timeliness, and good detection and recognition capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 The present invention discloses a flow chart of a small target detection and recognition algorithm for a forward-looking sonar image.

[0023] Figure 2 It is a principle block diagram of the constant false alarm algorithm detection of the present invention.

[0024] Figure 3 It is a schematic diagram of the principle of the HOG algorithm of the present invention.

[0025] Figure 4 It is a flow chart of the SVM algorithm of the present invention.

[0026] Figure 5 It is a flow chart of image preprocessing and image target classification and recognition of the present invention.

[0027] Figure 6 It is the sonar target detection and recognition diagram of the present invention. DETAILED DESCRIPTION

[0028] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The advantages and features of the present invention will become more apparent from the following description. It should be noted that the accompanying drawings are in very simplified form and in non-precise proportions, and are only used to conveniently and clearly assist in explaining the purpose of the embodiments of the present invention.

[0029] like Figure 1 As shown, the embodiment of the present invention specifically provides a small target detection and recognition algorithm for a forward-looking sonar image, comprising the following steps:

[0030] Step 1: Sonar beam domain data input;

[0031] Step 2: Detect the position of small targets through the constant false alarm algorithm;

[0032] The above-mentioned constant false alarm algorithm detection includes the following: constant false alarm detection is used to detect the possible target echo part in the echo signal. It consists of two major functional parts: background estimation and target detection. That is, after the data is input, background estimation is performed first, and then target detection is performed.

[0033] The input of CFAR detection is the raw data of beamforming output, and the output is the binary data of target detection. The functional block diagram of CFAR detection is as follows: Figure 2 As shown. The beam domain data is divided into several units by row (azimuth) or column (range). The unit size is determined by the target size, and the units are detected one by one. Protection units are set on the left and right sides of the detection unit. n units are selected on both sides as background unit groups. n is determined by the unit size and the working background. The maximum value of the average value of the background unit groups on both sides is calculated to calculate the detection threshold. Each element of the detection unit is compared with the detection threshold and binarized.

[0034] Among them, n is the number of units in the background unit group, x1~x n ,y1~y nis the background unit mean, X and Y are the sum of the mean values ​​of the background unit groups on both sides, Z is the maximum background unit group mean, T is the false alarm coefficient, S is the detection threshold, D is the value of each element in the detection unit, and the T value formula is as follows:

[0035]

[0036] Where, e is the natural exponent and P is the false alarm probability.

[0037] Step 3: Extract the target HOG features of the sonar image through the HOG algorithm; specifically include the following:

[0038] Image preprocessing, for the preprocessing of sonar images, that is, the detection window, first, when used with the classifier, the image training set needs to be resized. The size depends on the target size and the sonar image resolution, but it should not be too large, and the target should basically occupy the main part of the sample image. Secondly, gamma correction is performed on the sample image. Gamma correction can be understood as improving the image contrast effect of the dark or bright part of the image, which can effectively reduce the local shadows and intensity changes of the image.

[0039] Image feature extraction, the HOG algorithm is needed to extract features from sonar images. This algorithm uses statistical information of gradients to describe the appearance and shape of local targets. Extracting HOG features is mainly divided into three steps. The first step is to divide the sample image into units of several pixels, and then divide the gradient direction into 9 intervals. In each unit, the gradient direction of all pixels in each direction interval is histogram-statisticed to obtain a 9-dimensional feature vector; the second step is that each 4 adjacent units constitute a block, and the feature vectors in a block are connected in series to obtain a 36-dimensional feature vector and normalized; the third step is to scan the sample image with a block, and the scanning step size is a unit size. Finally, the features of all the blocks are connected in series to obtain a feature vector of a sample.

[0040] Step 4: Use the SVM algorithm to identify the type of detected targets and perform image classification.

[0041] The image classification uses the SVM algorithm. The basic model of the SVM algorithm is defined as a linear classifier with the largest interval in the feature space. Its learning strategy is to maximize the interval, which is ultimately converted into a solution to a convex quadratic programming problem, giving an optimal segmentation hypersurface.

[0042] The quadratic programming problem to be solved is:

[0043] maximize:

[0044] Constraints: 0≤α i ≤c(2)

[0045]

[0046] Decision plane intercept:

[0047] Linear kernel: k(X,Y)=X T Y (5)

[0048] Test polynomial:

[0049] If formula (6) holds, then Y O =1, otherwise Y O =-1.

[0050] Among them, (x, y) is the input binary label vector group, x is the feature vector, y is the label vector, N is the number of input samples, α is the optimal condition coefficient, c is a constant, B is the decision plane intercept, k(X,Y) is the linear kernel; X O Input feature vector for testing, Y O Output labels for the tests.

[0051] Based on the SVM algorithm, the target classification decision plane is obtained. First, the SVM kernel function is determined to be a linear kernel and the SVM algorithm parameters are configured; secondly, the optimal condition coefficient is iterated for the binary labeled HOG feature vector group and the decision plane intercept B is calculated to obtain the SVM model; finally, the obtained SVM model is tested and verified using other binary labeled HOG feature vector groups. The SVM process is as follows Figure 4 shown.

[0052] The above description is only a description of the preferred embodiments of the present invention, and is not intended to limit the scope of the present invention. Any changes or modifications made by a person skilled in the art in the field of the present invention based on the above disclosure shall fall within the scope of protection of the claims.

Claims

1. A small target detection and recognition algorithm for forward-looking sonar images, characterized in that: The steps include: Step 1: Sonar beam domain data input; Step 2: Detect the position of small targets through the constant false alarm algorithm; Step 3: Extract the target HOG features of the sonar image through the HOG algorithm; Step 4: Use the SVM algorithm to identify the type of detected targets and perform image classification.

2. The small target detection and recognition algorithm for forward-looking sonar images as claimed in claim 1, characterized in that: In the step 1, the sonar beam domain data is input to the constant false alarm algorithm detection, and the input is the original data output by the beamforming, and the output of the constant false alarm algorithm detection is the target detection binary data.

3. The small target detection and recognition algorithm for forward-looking sonar images as claimed in claim 1, characterized in that: In step 2, the constant false alarm algorithm detection is used to detect the target echo part in the echo signal, and the constant false alarm algorithm detection consists of two major functional parts: background estimation and target detection; that is, after data is input, background estimation is performed first, and then target detection is performed.

4. The small target detection and recognition algorithm for forward-looking sonar images as claimed in claim 1, characterized in that: Before extracting the target HOG features in step 3, the process also includes: image preprocessing, that is, first, when used with a classifier, the image training set needs to be resized, the size of which depends on the target size and the sonar image resolution, and the target only occupies the main part of the sample image; second, the sample image is gamma corrected.

5. The small target detection and recognition algorithm for forward-looking sonar images as claimed in claim 1, characterized in that: The target HOG feature extraction in step 3 specifically includes the following steps: The first step is to divide the sample image into several pixel units, and then divide the gradient direction into 9 intervals. In each unit, the gradient directions of all pixels in each direction interval are histogram-counted to obtain a 9-dimensional feature vector. In the second step, every 4 adjacent units form a block, and the feature vectors in a block are concatenated to obtain a 36-dimensional feature vector and normalized; The third step is to scan the sample image with blocks, with a scanning step size of one unit. Finally, the features of all blocks are connected in series to obtain a feature vector of a sample.

6. The small target detection and recognition algorithm for forward-looking sonar images as claimed in claim 1, characterized in that: The image classification is performed in step 4, which specifically includes the following steps: First, determine the SVM kernel function as a linear kernel and configure the SVM algorithm parameters; Secondly, the optimal condition coefficients are iterated for the binary labeled HOG feature vector group and the decision plane intercept B is calculated to obtain the SVM model; Finally, other binary labeled HOG feature vector groups are used to test and verify the obtained SVM model.

Images