A method for target category recognition based on sparse features

Abstract

The present invention relates to a sparse feature-based target classification method, belonging to the technical field of underwater target recognition and classification. Based on target echo signals received by a sonar array, the method utilizes the sparse representation of these echo signals in the dictionary domain as target features and employs a support vector machine classification algorithm for identification. This invention achieves excellent classification results without requiring deep learning, avoiding the need for typical samples for network training, and addressing the shortcomings of existing algorithms.

Description

Technical Field

[0001] The present invention belongs to the technical field of underwater target recognition and classification, and particularly relates to a classification and recognition method based on the sparse expression of target echo signals in a dictionary domain as target features and adopting a support vector machine classification algorithm. Background Art

[0002] Traditional classifiers work well when processing highly variable data. However, as the data becomes more complex, the requirements for feature extraction become higher, and the classifier's ability to represent target features and generalization decline. Due to the complexity and instability of underwater acoustic signals, traditional classification methods cannot effectively process the data and cannot accurately classify targets. This invention is essential to improve the efficiency and accuracy of target recognition and classification using target echo signals under low signal-to-noise ratios.

[0003] The paper "Research on Active Sonar Target Echo Recognition Based on Deep Learning" uses Bayesian regularization theory to derive network performance parameters from BP neural network training results. By adaptively adjusting the hyperparameters during network training, the paper optimizes network performance to achieve recognition and classification. However, the approach's approximation and generalization capabilities during neural network training are closely related to the representativeness of the learning samples, and selecting representative examples to form the training set is a challenging task. Summary of the Invention

[0004] Technical problems to be solved

[0005] In order to avoid the shortcomings of the prior art, the present invention provides a target category recognition method based on sparse features.

[0006] Technical Solution

[0007] A target category recognition method based on sparse features uses the sparse expression of received signals in the dictionary domain as target features and adopts a support vector machine classification algorithm for classification and recognition. The method is characterized by the following steps:

[0008] Step 1: Rotate the target with its geometric center as the origin, and divide the target rotation angle into N s Angle, expressed as Each angle θ i There is an echo signal x i , where i = 1, 2, ..., N s ; Construct the angle dictionary D1, which is (L*n)×N s The matrix of the matrix, the i-th (1≤i≤N s )

[0009] ψ i =[x1(1,θ i),…,x1(n,θ i ), x2(1, θ i ),…,x2(n,θ i ),…,x L (1, θ i ),…,x L (n, θ i )] T (1-1)

[0010] The angle dictionary of the target model Right now

[0011]

[0012] Similarly, the angle dictionary of different target models can be obtained;

[0013] Step 2: Merge the dictionaries of different models to form a joint dictionary D c ;

[0014] Step 3: Solve the signal x in the dictionary D c When the sparse expression α in

[0015]

[0016] Where γ represents the weight coefficient, which changes with noise. When different target echoes are substituted into the joint dictionary, different sparse expressions α can be solved. The sparse expressions α of targets of the same category are grouped into one category, which represents the sparse expression features of this category of targets in the joint dictionary.

[0017] Step 4: Feed the sparse expression feature labels into the SVM classifier for training. At this point, the received echo of the unknown target can be sparsely expressed in the joint dictionary, and the sparse expression feature β can be obtained using Equation (1-4). Finally, the sparse expression feature of the unknown target is input into the SVM classifier with category information to obtain the final classification result.

[0018] A further technical solution of the present invention: in step 2, when the combined dictionary D c , there are three goals, which can be specifically expressed as formula (1-3)

[0019]

[0020] A further technical solution of the present invention: Step 3 represents the sparse expression set of the echo of target 1 in the joint dictionary as α xi , the sparse expression set of target 2’s echo in the joint dictionary is represented as α yi , the sparse expression set of target 3’s echo in the joint dictionary is represented as α zi .

[0021] A computer system, characterized in that it includes: one or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned method.

[0022] A computer-readable storage medium is characterized by storing computer-executable instructions, which are used to implement the above method when executed.

[0023] Beneficial effects

[0024] Sparse representation classification is a new classification approach that has emerged in recent years. It was first used in face recognition and achieved good classification results. Traditional classifiers work well when dealing with highly diverse data, but as the data becomes more complex, the requirements for feature extraction become higher, and the classifier's ability to express target features and generalize decline.

[0025] This invention provides a method for target classification based on sparse features. This method uses the sparse representation of target echo signals received by a sonar array in the dictionary domain as target features and employs a support vector machine classification algorithm for identification. This method achieves excellent classification results without requiring deep learning, avoiding the need for typical samples for network training, and addressing the shortcomings of existing algorithms. Experimental verification has shown that this method achieves excellent results, has broad application prospects, and is readily available for immediate use. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The accompanying drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like parts throughout the drawings.

[0027] Figure 1 Flowchart of the invention.

[0028] Figure 2 The target dictionaries and coherence characteristic diagrams of the AUV model, torpedo model and mine model established in this invention are as follows: (a) target dictionary of the AUV model; (b) target dictionary of the torpedo model; (c) target dictionary of the mine model; (d) three-dimensional diagram of the coherence characteristics of the target dictionary of the AUV model; (e) three-dimensional diagram of the coherence characteristics of the target dictionary of the torpedo model; (f) three-dimensional diagram of the coherence characteristics of the mine model; (g) coherence characteristic contour diagram of the target dictionary of the AUV model; (h) coherence characteristic contour diagram of the target dictionary of the torpedo model; (i) coherence characteristic contour diagram of the mine model; (j) coherence characteristic side view of the AUV model; (k) coherence characteristic side view of the torpedo model; (l) coherence characteristic side view of the mine model.

[0029] Figure 3 (a) Three-dimensional map of dictionary coherence characteristics; (b) Contour map of dictionary coherence characteristics; (c) Contour map of dictionary coherence characteristics.

[0030] Figure 4 This is the principle diagram of the SVM dichotomy method in this invention.

[0031] Figure 5 It is the recognition rate of target categories based on sparse features of different emission signals in this invention. DETAILED DESCRIPTION

[0032] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only intended to illustrate the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0033] The basic idea of ​​the present invention is to establish a joint target dictionary, use the l1 norm minimization algorithm according to the echo data of different targets to extract the sparse expression features of different targets, send the sparse expression feature labels into the SVM classifier training, and then input the sparse expression features of the received echo of the target to be identified into the SVM classifier with category information to obtain the classification results.

[0034] Step 1: First construct the AUV model target dictionary D1, the torpedo model target dictionary D2 and the mine model target dictionary D3.

[0035] The AUV model is rotated with its geometric center as the origin, and the rotation angle is rotated 360° in steps of 1°. Each rotation angle direction corresponds to an element in the dictionary, that is, a dictionary D1 about angles is constructed, which is (L*n)×N s The matrix of the matrix, the i-th (1≤i≤N s )

[0036] ψ i =[x1(1,θ i ),…,x1(n,θ i ), x2(1, θ i ),…,x2(n,θ i ),…,x L (1, θ i ),…,x L (n, θ i )] T (1-1)

[0037] Then the dictionary of AUV model angles Right now

[0038]

[0039] Similarly, the torpedo model target dictionary D2 and the mine model target dictionary D3 are as follows:

[0040]

[0041]

[0042] Instructions attached Figure 2 It is a target dictionary based on double sparse random arrays and its coherence characteristics analysis, including AUV model, torpedo model and mine model.

[0043] As can be seen from Figure (1), when the rotation angle of the AUV model and the torpedo model changes, the time domain will be extended and shortened due to the large range of changes in the distribution of bright spots. However, due to the special shape of the mine model, the bright spots are basically distributed in the same range, and the time domain will be extended and compressed.

[0044] As can be seen from Figures (2), (3), and (4), the torpedo model has the narrowest main lobe and the lowest side lobe for the target dictionary coherence characteristic, followed by the AUV model. The mine model has a high side lobe for the coherence characteristic because the signal does not differ much with angle. Therefore, it is expected that the torpedo model has the best target dictionary rotation angle resolution.

[0045] Step 2: Combine the dictionaries of the three models to form a joint dictionary D c , specifically expressed as formula (1-5)

[0046]

[0047] Instructions attached Figure 3 It is a coherent characteristic diagram of the joint target dictionary of AUV, torpedo and mine models.

[0048] Step 3: Use the dictionary to extract sparse expression features of the three targets.

[0049] When the target receives the echo Assume that the target echo comes from the rotation angle θ i (represents the i-th rotation angle), then the l1 norm minimization method in convex optimization theory is used to solve it, that is, to solve

[0050]

[0051] The sparse expression set of the AUV model echo in the joint dictionary is expressed as α xi (1≤i≤N s ), the sparse expression set of the torpedo model echo in the joint dictionary is represented as α yi (1≤i≤N s), the sparse expression set of the mine model in the joint dictionary is represented as α zi (1≤i≤N s ).

[0052] Step 4: SVM classifier design

[0053] The principle of support vector machine (SVM) is to find a hyperplane in a sample space so that the blank areas on both sides of it are the largest. Figure 4 .

[0054] In the search for the maximum blank area hyperplane, w and b must meet the following requirements:

[0055] y i [(w·x i )+b]-1≥0

[0056] Under the constraints of the above formula, the problem of solving the optimal hyperplane can be transformed into the minimum value corresponding to the following function:

[0057]

[0058] The optimal solution of this function is expressed as:

[0059]

[0060] t i is the Lagrange multiplier, find the partial derivative with respect to w and b and set it equal to 0, so:

[0061]

[0062] For variables x and y, the mapping function to the new space is Then in the new space, the two correspond to and Its inner product is Let the function As the kernel function, after experimentation, we adopted the polynomial kernel function, namely

[0063] K(x i ,y i )=(v||x i -x j || a +r) b

[0064] Step 5: Feed the sparsely expressed feature labels into the SVM classifier for training. At this point, the received echo of the unknown target can be sparsely expressed in the joint dictionary to obtain the sparse expression feature β. The sparse expression of the unknown echo is then input into the trained SVM classifier to obtain the final classification result.

[0065] The targets are AUV models, torpedo models and mine models, and the emission signals are mainly bionic signals, that is, they contain first harmonic and second harmonic signals with a frequency band range of 35kHz-55kHz and 75kHz-95kHz. Other signals are used as references, including 60kHz-65kHz LFM signals, 60kHz-80kHz LFM signals, 35kHz-55kHz HFM signals, and 75kHz-95kHz HFM signals. The support vector machine training data for each signal uses a total of 3240 sparse expression features at -15dB, -5dB and 5dB signal-to-noise ratios. In order to verify the noise resistance and stability of the algorithm, the emission signal is changed, and the simulation results are shown in Table 1 and Table 2. Figure 5 shown.

[0066] From Table 1 and Figure 5 It can be seen that when the transmitted signal is a 60kHz-65kHz LFM signal, as the signal-to-noise ratio increases, the recognition rate increases, reaching 74.54% at SNR = -3dB; comparing the two LFM signals, as the bandwidth increases, the recognition rate reaches 80.09% at SNR = -3dB; the recognition rate of the first harmonic of the bionic signal reaches 86.11% at SNR = -3dB, and the recognition rate of the second harmonic of the bionic signal reaches 90.65% at SNR = -3dB; the recognition effect of the bionic signal is the best, and the recognition rate can reach 92.41% at SNR = -3dB, thus verifying the feasibility of the multi-target category recognition method based on sparse features, and the sparse feature target category recognition effect is best when the transmitted signal is a bionic signal.

[0067] Table 1 shows the invention Figure 5 The results are presented in the form of a table, making the results more intuitive

[0068]

[0069] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed in the present invention, and these modifications or replacements should all be included in the scope of protection of the present invention.

Claims

1. A target category recognition method based on sparse features, which uses the sparse expression of the received signal in the dictionary domain as the target feature and adopts the support vector machine classification algorithm for classification and recognition; Here are the steps: Step 1: Rotate the target with its geometric center as the origin, and divide the target rotation angle into N s Angle, expressed as Each angle θ i There is an echo signal x i , where i = 1, 2, ..., N s ; Construct the angle dictionary D1, which is (L*n)×N s The matrix of the matrix, the i-th (1≤i≤N s ) is listed as ψ i =[x1(1,θ i ),…,x1(n,θ i ), x2(1, θ i ),…,x2(n,θ i ),…,x L (1, θ i ),…,x L (n, θ i )] T (1-1) The angle dictionary of the target model Right now Similarly, the angle dictionary of different target models can be obtained; Step 2: Merge the dictionaries of different models to form a joint dictionary D c ; Step 3: Solve the signal x in the dictionary D c When the sparse expression α in Where γ represents the weight coefficient, which changes with noise. When different target echoes are substituted into the joint dictionary, different sparse expressions α can be solved. The sparse expressions α of targets of the same category are grouped into one category, which represents the sparse expression features of this category of targets in the joint dictionary. Step 4: The sparse expression feature label is sent to the SVM classifier for training; at this time, the received echo of the unknown target can be sparsely expressed in the joint dictionary, and the sparse expression feature β is obtained using formula (1-4). Finally, the sparse expression feature of the unknown target is input into the SVM classifier with category information to obtain the final classification result.

2. The target category recognition method based on sparse features according to claim 1 is characterized in that In step 2, when the joint dictionary D c , there are three goals, which can be specifically expressed as formula (1-3) 3. The target category recognition method based on sparse features according to claim 2 is characterized in that Step 3 represents the sparse expression set of the echo of target 1 in the joint dictionary as α xi , the sparse expression set of target 2’s echo in the joint dictionary is represented as α yi , the sparse expression set of target 3’s echo in the joint dictionary is represented as α zi .

4. A computer system, characterized in that include: One or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors are enabled to implement the method of claim 1.

5. A computer-readable storage medium, characterized in that Computer-executable instructions are stored, and when the instructions are executed, they are used to implement the method of claim 1.

Images