Target detection and recognition method based on one-dimensional high-resolution range profile

Abstract

The present invention proposes a target detection and recognition method based on one-dimensional high-resolution range image, which can realize the detection and recognition of the target based on the one-dimensional high-resolution range image signal with complex and diverse waveforms, and has good detection efficiency and high recognition accuracy. In view of the complex and diverse waveform characteristics of the one-dimensional high-resolution range image signal, the present invention first performs energy normalization preprocessing on the data signal; when the waveform of the data signal is relatively stable, the improved CA‑CFAR method is used for detection; and the detected target is preprocessed by power transformation to enhance the effect of weak scattering points of the target and reduce the shielding of weak scattering points by strong scattering points; finally, the improved Resnet18 is used for target classification and recognition. It can be applied to the detection and recognition of one-dimensional high-resolution range image signals, and can effectively improve the detection rate and recognition rate.

Description

Technical Field

[0001] The present invention relates to the field of radar signal processing technology, and in particular to a target detection and recognition method based on one-dimensional high-resolution range image. Background Art

[0002] High Resolution Range Profile (HRRP) signal is the vector and amplitude waveform of the target scattering point echo obtained by broadband radar projected on the radar line of sight. The range extended target and clutter in the one-dimensional high-resolution range profile are divided according to the intensity. Once there are a large number of non-targets with the same or even higher intensity as the target in the echo, it is difficult to eliminate the non-ship targets by simply using the target intensity information for ship target detection. Ship target detection is the premise and basis for ship target recognition. Once a false alarm occurs in the detection result, it will cause a meaningless recognition process, resulting in unnecessary waste of time and resources. Therefore, how to effectively utilize the target information provided by the one-dimensional high-range resolution radar echo in an efficient way is of great significance to the detection of one-dimensional high-resolution range profile.

[0003] One-dimensional high-resolution range image signal can effectively reflect the distribution of radar scattering cross-section of scatterers such as the hull, bow, and stern of the ship target on the radar line of sight, reflecting the relative position distribution of scattering points on the target, and contains rich information about the target. Ship targets have their own unique motion and physical characteristics. When the relative state (distance, azimuth, etc.) between the target and the radar changes, the echo of each scattering point will change, and the number of scattering points contained in each range unit may also change. Since the one-dimensional range image can be regarded as the superposition of the echoes of multiple scattering points in the range unit, the one-dimensional range image of the target will change accordingly. If the traditional method uses a single classifier for identification, the classifier is difficult to fit during the training process and cannot meet the requirements. At the same time, the target recognition model also faces the problems of target sample imbalance and different risk requirements of different types of target samples during the training process. Therefore, how to use limited one-dimensional high-resolution range image information is of great significance for the target classification and recognition of one-dimensional high-resolution range images. Summary of the invention

[0004] In view of this, the present invention proposes a target detection and recognition method based on one-dimensional high-resolution range image, which can realize the detection and recognition of the target based on the one-dimensional high-resolution range image signal with complex and diverse waveforms, and has good detection efficiency and high recognition accuracy.

[0005] To achieve the above object, the technical solution of the present invention is:

[0006] A target detection and recognition method based on one-dimensional high-resolution range image comprises the following steps:

[0007] Step 1: Acquire the signal data of the one-dimensional high-resolution range image, mark the target on the one-dimensional high-resolution range image data, and extract the one-dimensional target signal by reading line by line;

[0008] Step 2: Perform energy normalization preprocessing on the one-dimensional high-resolution range image target signal;

[0009] Step 3: Perform CA-CFAR detection on the processed target signal data, and detect the targets close to the left and right sides of the one-dimensional data for the protection unit and training unit of the CA-CFAR method;

[0010] Step 4: Perform power transformation on the detected target signal and classify it;

[0011] Step 5: According to the characteristics of the one-dimensional high-resolution range image target signal, the Resnet18 network structure is improved, and target classification and recognition training and testing are performed on the improved network.

[0012] Among them, in step 3, the specific steps of CA-CFAR detection are as follows:

[0013] Set the detection parameters, set the number of protection units pro_N = 100, the number of training units N = 128, and the false alarm rate α = 10^(-1.5);

[0014] The one-dimensional high-resolution range image data to be detected is filled with N / 2+pro_N / 2 on both sides and filled with zeros. The total length of the data is 1, and the total length after filling is 1+N+pro_N, that is, L. The training unit data on the left and right sides are calculated by "Z" to obtain the Z value.

[0015] After obtaining the Z value, the detection value T is obtained by performing a dot product calculation with the false alarm rate;

[0016] Compare the detected value T with the target true value D to obtain the detection result.

[0017] The strategy for calculating "Z" is as follows:

[0018] In the interval (N+pro_N, LN-pro_N),

[0019] In the interval (N / 2+pro_N / 2, N+pro_N),

[0020] In the interval (LN-pro_N, LN / 2-pro_N / 2),

[0021] Among them, i represents the position of the target true value D in the data, N / 2+pro_N / 2<i<LN / 2-pro_N / 2.

[0022] Among them, in the step 5, under the improved network structure of Resnet18, the data is divided into a training set and a test set at a ratio of 3:1. During the training process, the batch is set to 16. For 16 groups of 1*512 target signals, the final average pooling downsampling multiple is reduced to 4 times; one layer of full connection is changed to two layers of convolution connected to each other; and the optimal training model is tested on the test set to obtain the classification results of the two types of ships.

[0023] Among them, in step 2, energy normalization is specifically as follows:

[0024] The total energy μ of the target signal is normalized for each pixel. After normalization, the total energy of the target signal is 1. The formula for energy normalization is as follows:

[0025]

[0026] Among them, μ refers to the total energy of the target signal, δ represents the value of a pixel point in the target signal, and θ=1, 2, 3...μ.

[0027] Beneficial Effects

[0028] 1. Aiming at the complex and diverse waveform characteristics of one-dimensional high-resolution range image signals, the present invention first performs energy normalization preprocessing on the data signal; when the waveform of the data signal is relatively stable, the improved CA-CFAR method is used for detection; and the detected target is preprocessed by power transformation to enhance the effect of weak scattering points of the target and reduce the shielding of weak scattering points by strong scattering points; finally, the improved Resnet18 is used for target classification and recognition. It can be applied to the detection and recognition of one-dimensional high-resolution range image signals, and can effectively improve the detection rate and recognition rate.

[0029] 2. The present invention improves the CA-CFAR method, specifically by filling the same blanks on the left and right sides of the one-dimensional data; and calculating "Z" for the edge data using the mean of the non-filled data. In terms of target detection in one-dimensional high-resolution range images, the improved CA-CFAR can achieve a good detection rate regardless of whether the target is relatively simple around the target, or the environment is more complex when the land and sea are combined, or the number of targets is relatively large and mixed.

[0030] 3. According to the characteristics of the one-dimensional high-resolution range image target signal, the present invention improves the Resnet18 network structure, divides the data into a training set and a test set at a ratio of 3:1, and sets the batch to 16 during the training process. The 16 groups of 1*512 target signals should have gone through 17 convolution layers, 2 pooling layers, and 1 fully connected layer in the Resnet18 network. However, because the one-dimensional high-resolution range image data length is relatively small, the accuracy requirement is relatively high, and only two types of ships need to be classified, the downsampling multiple of the final average value pooling is reduced to 4 times. At the same time, in order to improve the learning ability of the model, one layer of full connection is changed to two layers of convolution connected by full connection layers. The optimal training model is tested on the test set to obtain the classification results of the two types of ships. When the test data volume is only 40, the recognition rate can still reach 87.5%. . BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Schematic diagram of the method of the present invention.

[0032] Figure 2 This is the CA-CFAR detection flow chart of the present invention.

[0033] Figure 3 This is a schematic diagram of the improved CA-CFAR detection results of the present invention.

[0034] Figure 4 This is the improved network structure diagram of Resnet18 of the present invention. DETAILED DESCRIPTION

[0035] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0036] The present invention provides a method for target detection and recognition based on one-dimensional high-resolution range image. In view of the complex and diverse waveform characteristics of one-dimensional high-resolution range image signals, the data signal is first preprocessed by energy normalization; when the waveform of the data signal is relatively stable, the improved CA-CFAR method is used for detection; and the detected target is preprocessed by power transformation to enhance the effect of weak scattering points of the target and reduce the shielding of weak scattering points by strong scattering points; finally, the improved Resnet18 is used for target classification and recognition. The overall flow chart is as follows: Figure 1 As shown, the following steps are included:

[0037] Step 1: Obtain signal data of a one-dimensional high-resolution range image, mark the one-dimensional high-resolution range image data with targets, and extract the one-dimensional target signal by reading line by line. In this embodiment, the one-dimensional target signal is a one-dimensional target signal with a azimuth of 400 meters and a range image of 2 meters.

[0038] Step 2: Perform energy normalization preprocessing on the one-dimensional high-resolution range image target signal.

[0039] Energy normalization is as follows:

[0040] The total energy μ of the target signal is normalized for each pixel, and the total energy of the target signal after normalization is 1. That is, norm(δ, 2), and the formula for energy normalization is as follows:

[0041]

[0042] Among them, μ refers to the total energy of the target signal, δ represents the value of a pixel point in the target signal, and θ=1, 2, 3...μ.

[0043] Step 3: Perform CA-CFAR detection on the processed target signal data, and focus on the impact of the protection unit and training unit of the CA-CFAR method on the detection of targets close to the left and right sides in the one-dimensional data.

[0044] The present invention improves the CA-CFAR method by: 1. adding blank padding to one-dimensional data; 2. using the mean of non-filled data to calculate "Z" for the edge data. The CA-CFAR detection process of the present invention is as follows: Figure 2 shown.

[0045] In terms of target detection of one-dimensional high-resolution range images: Whether the target is relatively simple, or the environment is more complex, or the number of targets is relatively large and mixed, the improved CA-CFAR can achieve a good detection rate (the comparison results are shown in the figure). Figure 3 ). The specific steps are as follows:

[0046] 1) Set the detection parameters, set the number of protection units pro_N = 100, the number of training units N = 128 and the false alarm rate α = 10^(-1.5);

[0047] 2) The traditional CA-CFAR algorithm discards the left and right sides when processing one-dimensional signal data.

[0048] For data of size N / 2+pro_N / 2, this method adopts the method of filling with zeros. The left and right sides of the one-dimensional high-resolution range image data to be detected are filled with N / 2+pro_N / 2 and zeros. Assuming that the total length of the data is l, the total length after filling is l+N+pro_N, that is, L. The training unit data on the left and right sides are used to calculate "Z". The strategy for calculating "Z" is as follows (where i represents the position of the target true value D in the data, N / 2+pro_N / 2 <i<L-N / 2-pro_N / 2):

[0049] ①In the interval (N+pro_N, LN-pro_N)

[0050]

[0051] ②In the interval (N / 2+pro_N / 2, N+pro_N)

[0052]

[0053] ③In the interval (LN-pro_N, LN / 2-pro_N / 2)

[0054]

[0055] 3) After obtaining the Z value, the detection value T is obtained by performing a dot product calculation with the false alarm rate.

[0056]

[0057] 4) Compare the detected value T with the target true value D to obtain the detection result.

[0058] Step 4: Perform power transformation on the detected 1*512 target signal and classify it.

[0059] The purpose of power transformation is to enhance the effect of weak scattering points and reduce the shielding of weak scattering points by strong scattering points. v (0<v<1), y is the value after power transformation, x is the value of the signal to be processed, v is the power transformation coefficient, 0<v<1, the higher the value, the more obvious the effect.

[0060] Step 5: According to the characteristics of the one-dimensional high-resolution range image target signal, the Resnet18 network structure is improved, and the target classification and recognition training and testing are performed on the improved network. Figure 4 As shown in the figure, the data is divided into training set and test set in a ratio of 3:1. During the training process, the batch size is set to 16. The 16 groups of 1*512 target signals should have gone through 17 convolution layers, 2 pooling layers, and 1 fully connected layer in the Resnet18 network. However, because the length of the one-dimensional high-resolution distance image data is relatively small, the accuracy requirement is relatively high, and only two types of ships need to be classified, the downsampling multiple of the final average pooling is reduced to 4 times. At the same time, in order to improve the learning ability of the model, one fully connected layer is changed to two fully connected layers of convolution. The optimal training model is tested on the test set to obtain the classification results of the two types of ships, as shown in Table 1. When the test data volume is only 40, the recognition rate can still reach 87.5%.

[0061] Table 1 Classification results of two types of ships

[0062]

[0063] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A target detection and recognition method based on one-dimensional high-resolution range image, It is characterized in that The steps include: Step 1: Acquire the signal data of the one-dimensional high-resolution range image, mark the target on the one-dimensional high-resolution range image data, and extract the one-dimensional target signal by reading line by line; Step 2: Perform energy normalization preprocessing on the one-dimensional high-resolution range image target signal; Step 3: Perform CA-CFAR detection on the processed target signal data, and detect the targets close to the left and right sides of the one-dimensional data for the protection unit and training unit of the CA-CFAR method; Step 4: Perform power transformation on the detected target signal and classify it; Step 5: According to the characteristics of the one-dimensional high-resolution range image target signal, the Resnet18 network structure is improved, and target classification and recognition training and testing are performed on the improved network; In step 3, the specific steps of CA-CFAR detection are as follows: Set the detection parameters, set the number of protection units pro_N = 100, the number of training units N = 128, and the false alarm rate α = 10^(-1.5); The one-dimensional high-resolution range image data to be detected is filled with N / 2+pro_N / 2 on both sides and zeros. The total length of the data is l, and the total length after filling is l+N+pro_N, that is, L. The training unit data on the left and right sides are calculated by Z to obtain the Z value. After obtaining the Z value, the detection value T is obtained by performing a dot product calculation with the false alarm rate; Compare the detected value T with the target true value D to obtain the detection result; The strategy for Z calculation is as follows: In the interval (N+pro_N, LN-pro_N), In the interval (N / 2+pro_N / 2, N+pro_N), In the interval (LN-pro_N, LN / 2-pro_N / 2), Among them, i represents the position of the target true value D in the data, N / 2+pro_N / 2 <i<L-N / 2-pro_N / 2。 2. The method according to claim 1, It is characterized in that In step 5, under the improved network structure of Resnet18, the data is divided into a training set and a test set at a ratio of 3:

1. During the training process, the batch is set to 16. For 16 groups of 1*512 target signals, the downsampling multiple of the final average pooling is reduced to 4 times; one layer of full connection is changed to two layers of convolution connected to each other; and the optimal training model is tested on the test set to obtain the classification results of the two types of ships.

3. The method according to claim 1 or 2, It is characterized in that In step 2, energy normalization is specifically as follows: The total energy μ of the target signal is normalized for each pixel. After normalization, the total energy of the target signal is 1. The formula for energy normalization is as follows: Among them, μ refers to the total energy of the target signal, δ represents the value of a pixel point in the target signal, and θ = 1, 2, 3…μ.

Images