A radar waveform classification method and system based on effective area recognition

Abstract

The application provides a radar waveform classification method and system based on effective area identification, and the method comprises the following steps: inputting a radar emitter time-frequency image into a preset SSD network trained in advance to obtain an effective area where a single intra-pulse signal is located and a corresponding radar waveform category output by the preset SSD network; the preset SSD network is trained by taking multiple radar emitter time-frequency images as training samples, taking the effective area where the single intra-pulse signal is located in each radar emitter time-frequency image and the corresponding radar waveform category as labels. The application identifies the effective area where the intra-pulse signal of the radar emitter time-frequency image is located by using the SSD network, and only classifies the category of the waveform in the effective area in the subsequent process, thereby avoiding the influence of signal characteristics in the invalid area on identification, and effectively improving the identification accuracy of the radar waveform.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image processing, more particularly, to a radar waveform classification method and system based on effective area recognition. BACKGROUND

[0002] With the wide use of radar in military and civilian fields, radar countermeasure reconnaissance has become a major reconnaissance form in electromagnetic spectrum warfare. Radar waveform classification, as an important part of radar countermeasure information processing, is one of the key technologies of radar countermeasure reconnaissance system, and its result is an important basis for judging the threat level of enemy weapons.

[0003] Currently, in the field of radar waveform classification, the conventional image classification method is mostly used, the radar emitter signal is converted into a time-frequency image, the CNN is used to extract the intra-pulse signal features of the whole image, and the waveform classification is completed. However, in the actual space, electromagnetic signals are dense, especially new system radar signals are difficult to intercept, and it cannot be guaranteed that the object to be identified is a complete intra-pulse signal. The intra-pulse signal may have a defect. If the features of the whole time-frequency image are directly extracted, the identification area is easy to be disturbed by invalid signal features, the identification area is not focused enough, the image features to be identified are greatly increased, and then the accuracy of radar waveform classification is affected, which may have a major impact on the subsequent decision-making and deployment of commanders. Therefore, the above problems have greatly restricted the performance of radar waveform classification, and further research on the actual space radar waveform classification method is needed to improve the identification accuracy. SUMMARY

[0004] The present application provides a radar waveform classification method and system based on effective area recognition to solve the technical problems in the prior art, which can improve the identification accuracy of radar waveform.

[0005] According to the first aspect of the present application, a radar waveform classification method based on effective area recognition is provided, which comprises: inputting a radar emitter time-frequency image into a preset SSD network trained, obtaining an effective area where a single intra-pulse signal is located and a corresponding radar waveform category output by the preset SSD network; the preset SSD network is trained with multiple radar emitter time-frequency images as training samples, and the effective area where a single intra-pulse signal is located in each radar emitter time-frequency image and the corresponding radar waveform category as labels.

[0006] On the basis of the above technical solution, the present application can also be improved as follows.

[0007] Optionally, the training samples and labels of the SSD network are obtained in the following way: a plurality of signal-to-noise ratios are determined, a plurality of radar waveform categories corresponding to the time-frequency images under each signal-to-noise ratio are collected; the effective area where a single intra-pulse signal is located in each time-frequency image is labeled, and the radar waveform category of the single intra-pulse signal is labeled.

[0008] Optionally, when there are multiple cycles of the intra-pulse signal contained in the time-frequency image, an effective area in which the intra-pulse signal most completely displayed in the time-frequency image is framed to obtain an effective area target frame of the intra-pulse signal, and a radar waveform category of the effective area target frame is marked.

[0009] Optionally, the preset SSD network comprises a backbone network, a detection network and an output network; the backbone network extracts multiple effective feature layers of the time-frequency image in a feature pyramid manner; the detection network is configured to perform twice convolution on each effective feature layer to obtain adjustment information of multiple prior frames and a radar waveform category corresponding to each prior frame; and the output network is configured to obtain a prior frame with the highest matching degree from the multiple prior frames as a predicted frame output.

[0010] Optionally, the output network is configured to obtain a prior frame with the highest matching degree from the multiple prior frames as a predicted frame output, comprising: calculating an area intersection over union (IOU) between each prior frame and the effective area target frame for any time-frequency image; discarding a prior frame with an area IOU less than an IOU threshold with the effective area target frame, and retaining a prior frame with an area IOU greater than or equal to the IOU threshold with the effective area target frame; obtaining a radar waveform category corresponding to each retained prior frame; and obtaining a prior frame with the highest confidence in the radar waveform category as the predicted frame output.

[0011] Optionally, the obtaining of the radar waveform category corresponding to each retained prior frame comprises: performing threshold judgment on all prior frames corresponding to each prediction category of the radar waveform for any prior frame, and obtaining a prior frame with a confidence greater than a threshold.

[0012] Optionally, the obtaining of the prior frame with the highest confidence in the radar waveform category as the predicted frame output comprises: for the multiple retained prior frames and the radar waveform category corresponding to each prior frame, using a non-maximum suppression algorithm to obtain a prior frame with the highest confidence as the predicted frame output.

[0013] Optionally, the method further comprises: comparing the predicted frame and the radar waveform category of the intra-pulse signal in each output time-frequency image with the marked effective area and radar waveform category, and calculating an identification accuracy of the preset SSD network according to a comparison result; repeatedly training the preset SSD network to improve the identification accuracy of the preset SSD network until the SSD network converges.

[0014] According to a second aspect of the present application, a radar waveform classification system based on effective area recognition is provided, comprising: a training module configured to train an SSD network using a plurality of radar emitter time-frequency images as training samples, and using an effective area of a single intra-pulse signal in each radar emitter time-frequency image and a corresponding radar waveform category as labels; and an acquisition module configured to input a radar emitter time-frequency image into the trained preset SSD network, and acquire an effective area of a single intra-pulse signal and a corresponding radar waveform category output by the preset SSD network.

[0015] According to a third aspect of the present application, an electronic device is provided, comprising a memory and a processor configured to implement the steps of the radar waveform classification method based on effective area recognition when executing a computer management program stored in the memory.

[0016] According to a fourth aspect of the present application, a computer readable storage medium is provided, having a computer management program stored thereon, the computer management program being configured to implement the steps of the radar waveform classification method based on effective area recognition when executed by a processor.

[0017] The radar waveform classification method and system based on effective area recognition provided by the present application use an SSD network to identify an effective area of an intra-pulse signal of a radar emitter time-frequency image, and subsequently identify only the category of the waveform in the effective area, thereby improving the identification accuracy of the radar waveform. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 A radar waveform classification method based on effective area recognition is provided by the present application;

[0019] FIG. 2(a) is a schematic diagram of an effective area of a single intra-pulse signal labeled in a time-frequency image;

[0020] FIG. 2(b) is a schematic diagram of radar waveform prediction classification results in a time-frequency image;

[0021] Figure 3 A schematic diagram of time-frequency image size adjustment is provided;

[0022] Figure 4 A radar waveform classification method based on effective area recognition is provided by the present application;

[0023] Figure 5 A radar waveform classification system based on effective area recognition is provided by the present application;

[0024] Figure 6 A hardware structure schematic diagram of a possible electronic device is provided by the present application;

[0025] Figure 7 A possible hardware structure schematic diagram of a computer readable storage medium provided by the present application. DETAILED DESCRIPTION

[0026] The specific embodiments of the present application are described in further detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present application, but are not used to limit the scope of the present application.

[0027] Figure 1 A method flowchart based on effective area recognition provided by the present application is shown in Figure 1 The method comprises: inputting a radar emitter time-frequency image into a preset SSD network trained, acquiring an effective area where a single intra-pulse signal is located and a corresponding radar waveform category output by the preset SSD network; the preset SSD network is trained with multiple radar emitter time-frequency images as training samples, and the effective area where the single intra-pulse signal is located in each radar emitter time-frequency image and the corresponding radar waveform category are used as labels.

[0028] It can be understood that, based on the defects in the background art, the embodiments of the present application propose a method which can identify the effective area of the intra-pulse signal in the radar emitter time-frequency image, and then identify the corresponding radar waveform category according to the intra-pulse signal in the identified effective area, thereby improving the identification accuracy of the radar waveform compared with the existing identification of the radar waveform category of the entire time-frequency image.

[0029] Specifically, the GNU Radio software radio development platform can realize the sending and receiving of the software defined radio, the hardware circuit structure is changed by using the software configuration update, and the repeated design of the hardware is avoided. After the HackRF receiver receives the radar radiation source signal, the radar radiation source signal is converted into a time-frequency image by using a time-frequency conversion algorithm. A single intra-pulse signal in the time-frequency image is framed out by using a visual image calibration tool to form an effective area, and the frame is named as a real frame. First, the real frame and the real label are labeled for each time-frequency image to generate a set of training test data set, then the backbone network of the SSD network (the SSD network is a target detection network) is used to extract multiple effective feature layers of the time-frequency image, and then the detection network of the SSD network has the characteristics of detecting the position of the target, and the effective area of the time-frequency image is detected. Since a large number of prior frames on the effective feature layer will generate a large number of matching prediction frames, it is necessary to use a non-maximum suppression algorithm to retain the prediction frames with low overlap and high confidence, so as to automatically and accurately identify the area where the intra-pulse signal is located. Finally, the evaluation mechanism in image classification is adopted, and the class with the highest prediction classification confidence is selected as the prediction class of the time-frequency image. The SSD network is trained to converge, and finally has the ability to accurately identify the radar waveform in the effective area of the time-frequency image, avoids the influence of the signal features in the irrelevant area on the waveform classification, and effectively improves the waveform classification accuracy.

[0030] The embodiment of the present application uses the SSD network to identify the effective area of the intra-pulse signal of the radar radiation source time-frequency image, and only identifies the category of the waveform in the effective area subsequently, thereby improving the identification accuracy of the radar waveform.

[0031] In a possible embodiment, the training samples and labels of the SSD network are obtained in the following manner: a plurality of signal-to-noise ratios are determined, and a plurality of time-frequency images corresponding to radar waveform categories under each signal-to-noise ratio are collected; the effective area of a single intra-pulse signal in each time-frequency image is labeled, and the radar waveform category of the single intra-pulse signal is labeled.

[0032] The specific process of training the SSD network is as follows: first, collect the training samples, and for different signal-to-noise ratios, collect time-frequency images of a plurality of radar waveforms under each signal-to-noise ratio. The signal in the single intra-pulse period in the time-frequency image is labeled by using a real frame, the area in the frame is an effective area, and the area outside the frame is an invalid area, that is, the effective area of the single intra-pulse signal in the time-frequency image is labeled, and the radar waveform category corresponding to the intra-pulse signal in the effective area is labeled. A large number of time-frequency images are used as training samples, and the effective area of the intra-pulse signal in each time-frequency image and the radar waveform category are used as labels to train the SSD network.

[0033] In a possible embodiment, when there are multiple cycles of the intra-pulse signal contained in the time-frequency image, the effective area in which the intra-pulse signal is most completely displayed in the time-frequency image is framed to obtain an effective area target frame of the intra-pulse signal, and the radar waveform category of the effective area target frame is marked.

[0034] It should be noted that there can be multiple cycles of the intra-pulse signal in each time-frequency image, and the embodiment of the present application only frames and marks the effective area in which the most complete intra-pulse signal is located, which is referred to as an effective area target frame below, and the radar waveform category of the effective area target frame is marked. It should be noted that the radar waveform category is identified according to the complete intra-pulse signal, and the accuracy of identification is higher.

[0035] The input image size of the SSD network is required to be 300*300, and the size of the time-frequency image is less than the input requirement, so the image size needs to be adjusted. To avoid image stretching leading to distortion of the time-frequency image, the time-frequency image should be filled with gray bars to meet the size requirement, and the adjusted size of the time-frequency image can be seen from FIG. 2. Figure 3

[0036] In a possible embodiment, the preset SSD network includes a backbone network, a detection network, and an output network; the backbone network extracts multiple effective feature layers of the time-frequency image in a feature pyramid manner; the detection network is configured to perform twice convolution on each effective feature layer to obtain adjustment information of multiple prior frames and a radar waveform category corresponding to each prior frame; and the output network is configured to obtain, from the multiple prior frames, a prior frame with the highest matching degree as a predicted frame output.

[0037] It can be understood that each time-frequency image in the training sample is input into the SSD network, and the SSD network includes a backbone network, a detection network, and an output network. The backbone network extracts 6 effective feature layers of the time-frequency image in a feature pyramid manner, and then the detection network performs twice convolution on each effective feature layer to obtain each prior frame (which can be understood as multiple target frames detected by the detection network) and a prediction of all signal categories of each prior frame. The output network adopts a non-maximum suppression algorithm to obtain a predicted frame with lower overlap and higher confidence from all prior frames.

[0038] In a possible embodiment, the output network is configured to obtain, from the multiple prior frames, a prior frame with the highest matching degree as a predicted frame output, and includes the following steps: for any time-frequency image, calculating an area intersection over union (IOU) between each prior frame and the effective area target frame; discarding a prior frame with an area IOU less than an IOU threshold with the effective area target frame, and retaining a prior frame with an area IOU greater than or equal to the IOU threshold with the effective area target frame; obtaining a radar waveform category corresponding to each retained prior frame; and obtaining a prior frame with the highest confidence of the radar waveform category as a predicted frame output.​

[0039] It can be understood that the specific method of the output network obtaining the prior box with the highest matching degree from all prior boxes as the output prediction box is that the true box (i.e., the labeled valid area target box) on each time-frequency image in the training sample and all prior boxes are iteratively calculated for IOU, the prior boxes lower than the IOU threshold are discarded, and the prior boxes greater than or equal to the IOU threshold are retained. For each retained prior box, the radar waveform class corresponding to each prior box is obtained.

[0040] wherein, IOU is the area intersection over union of the prior box and the true box, and the calculation formula is as follows,

[0041]

[0042] In the formula, S default represents the area of the prior box, S ground represents the area of the true box.

[0043] In a possible embodiment, obtaining the radar waveform class corresponding to each retained prior box includes: for each prediction class corresponding to any prior box, performing threshold judgment on all prior boxes corresponding to each prediction class to obtain the prior box with a confidence greater than a threshold.

[0044] It can be understood that for each retained prior box, there are multiple radar waveform prediction classes, and the confidence of each radar waveform prediction class is scored to obtain the prior box with a confidence score higher than a threshold in each prediction class, so that the number of prior boxes is further reduced.

[0045] In a possible embodiment, obtaining the prior box with the highest confidence of the radar waveform class as the output prediction box includes: for the retained multiple prior boxes and the radar waveform class corresponding to each prior box, using a non-maximum suppression algorithm to obtain the prior box with the highest confidence as the output prediction box.

[0046] It can be understood that for each retained prior box and the radar waveform class corresponding to each prior box, a non-maximum suppression algorithm is used to retain the prior box with a lower overlap and a higher confidence, and the prior box with the highest prediction class confidence is obtained as the output prediction box.

[0047] In a possible embodiment, the method further includes: comparing the prediction box and the radar waveform class of the intrapulse signal in each output time-frequency image with the labeled valid area and the radar waveform class, calculating the recognition accuracy of a preset SSD network according to the comparison result, repeatedly training the preset SSD network, improving the recognition accuracy of the preset SSD network, and stopping until the SSD network converges.

[0048] Specifically, for the training sample, the prediction result and the true result are loss calculated, and through a large number of training, the loss value is gradually reduced, and finally a more matched corresponding relationship is found.

[0049] Wherein, the total loss function loss is composed of positioning loss and confidence loss two parts:

[0050] L(x,c,p,g)=(ωL loc (x,p,g)+L conf (x,c)) / N;

[0051] In the formula, c represents the confidence of each signal category, p represents the prediction frame, g represents the true frame, ω represents the weight term, and the value is 1, and N is the number of matched prior frames. L loc (g) is the positioning loss of the prediction frame and the true frame, and SmoothL1 loss is adopted. conf (g) is the classification confidence loss, and the Softmax function is adopted, and its expression is as follows:

[0052]

[0053] In the formula, the first term is the positive sample confidence loss, and the second term is the negative sample confidence loss. x ij Indicates that the ith prior frame matches the jth true frame of the signal category m, and the value is 1, indicating matching, and the value is 0, indicating not matching.

[0054] After training the SSD network, the test set is used to test the trained SSD network, first, the effective feature layer of each time-frequency image in the test set is extracted, the position of each prior frame and the prediction of all signal categories are obtained. Through stacking and activation function, the position information of all prior frames and the confidence of the predicted category are obtained. Then the format of the prediction frame information is converted into the format of the true frame information with center point, frame width and frame height, then the threshold value of all prediction frames corresponding to each prediction category is judged, and the prediction frame with confidence greater than the threshold value is taken out, and the non-maximum suppression algorithm is used to reserve the prediction frames with low overlap and high confidence between the prediction frames. The reserved prediction frames are sorted according to the confidence of the prediction category from high to low, and the prediction frame with the highest confidence is taken out, so as to identify the effective area on the test set time-frequency image, and complete the prediction frame and confidence drawing.

[0055] According to the characteristics of radar waveform classification, the evaluation index is optimized, and the evaluation mechanism in image classification is adopted. In each time-frequency image of the test set, the prediction box with the highest prediction classification confidence and the corresponding prediction category are selected as the final prediction result of the time-frequency image. By comparing the prediction category and the real label, the recognition accuracy of a certain waveform under a certain signal-to-noise ratio is the ratio of the number of time-frequency images of the waveform correctly predicted to the number of all time-frequency images of the waveform. The average of the recognition accuracies of various waveforms under the same signal-to-noise ratio is the total recognition accuracy under the signal-to-noise ratio.

[0056] The radar waveform classification method based on effective area recognition provided by the application is introduced below with a specific example, which can be seen from Figure 4 , and mainly includes the following steps:

[0057] (1) The radar waveform is received through the software radio platform, and the Choi-Williams distribution is used to convert the radar waveform into a time-frequency image and saved as a.jpg format. The Choi-Williams distribution can effectively reduce the cross terms generated in the conversion process through the scale factor. The hardware part of the software radio platform is composed of HackRF and PC, and the software part is composed of GNURadio and Linux system.

[0058] (2) The single intra-pulse period signal frame in the time-frequency image is selected by using the label tool labelimg, and the frame is the real frame. The area in the frame is the effective area, and the area outside the frame is the invalid area, as shown in Fig. 2(a). At the same time, the real label (real category of radar waveform signal) is labeled for the signal in the frame, and the.xml file containing the real label and real frame coordinate information is generated. The label index file Class.txt is made, and the background category is 0, and the signal category starts from 1 and is arranged in order.

[0059] (3) The input image size of the SSD network is required to be 300*300, while the size of the time-frequency image is 244*256, which is smaller than the input requirement, so the image size needs to be adjusted. Simple image stretching can easily lead to image distortion, therefore, the padding method should be used to complete the size adjustment. First, the scale factor of the time-frequency image and the input image size requirement is calculated, and the width and height of the time-frequency image are enlarged by the same scale factor, then a 300*300 gray-white image is generated, and the time-frequency image is pasted to the corresponding coordinate position of the gray-white image through calculation, so as to meet the requirement, as shown in Figure 3 .

[0060] (4) The SSD network includes a backbone network, a detection network, and an output network. The backbone network selects VGG16, and replaces the fully connected layers FC6 and FC7 with convolutional layers. The backbone network adopts a feature pyramid method to perform multiple convolutions and pooling on the time-frequency image, and extracts 6 effective feature layers. The number of grids on each feature layer is different, and is 1444, 361, 100, 25, 9, and 1 respectively. At the same time, there are 4-6 prior boxes on each grid. Therefore, the total number of prior boxes on each feature layer is the number of grids multiplied by the number of prior boxes on each grid, thereby meeting the recognition needs of different sizes of targets in the image. Then the detection network performs two convolutions on each feature layer to obtain the adjustment of each prior box and the prediction of all signal categories corresponding to the prior box. Finally, the output network uses a non-maximum suppression algorithm to obtain the prediction box with lower overlap and higher confidence.

[0061] (5) Training: The training set is composed of time-frequency images under 7 signal-to-noise ratios, and the signal-to-noise ratio is from -9dB to 9dB with a step of 3dB. There are 200 time-frequency images for each signal under each signal-to-noise ratio. For example, there are 6 radar waveforms, BPSK, LFM, P1, P2, P3, and P4. The training set has a total of 8400 time-frequency images. A certain proportion of the data set can be randomly divided into a validation set. Training is performed through the Tensorflow-GPU deep learning framework. The number of epoch iterations is 120. The learning rate is large in the first 60 times for fast learning of signal characteristics, and the learning rate is small in the last 60 times for stable small-step learning of signal characteristics, so that the network converges gradually to the global optimum. In the training process, in order to speed up the training and improve the training effect, breakpoint resume training can be adopted, and the initial weight file is loaded, for example, the.h5 weight file of the completed training of the same type of data set.

[0062] First, the IOU of the real box and all prior boxes on the training set time-frequency image is calculated. The prior boxes with an IOU lower than the threshold are discarded, and the corresponding relationship between each prior box and the real box with the maximum overlap degree is preserved. The IOU threshold is 0.5.

[0063] The IOU is the area intersection ratio of the prior box and the real box, and the calculation formula is as follows:

[0064]

[0065] In the formula, S default represents the area of the prior box, and S ground represents the area of the real box.

[0066] Then, the information of the real box is converted into the format of the prediction result through encoding. The training set is predicted, and the loss is calculated between the prediction result and the real result. Through a large number of training, the loss value is gradually reduced, and finally a more matched corresponding relationship is found.

[0067] The total loss function loss is composed of two parts: positioning loss and confidence loss:

[0068] L(x, c, p, g) = (ωL loc (x, p, g) + L conf (x, c)) / N.

[0069] In the formula, c represents the confidence of each signal category, p represents the predicted box, g represents the real box, ω represents the weight term, which is 1, and N is the number of matched prior boxes. L loc (g) is the positioning loss of the predicted box and the real box, which adopts SmoothL1 loss. conf (g) is the classification confidence loss, which adopts Softmax function, and its expression is as follows:

[0070]

[0071] In the formula, the first term is the positive sample confidence loss, and the second term is the negative sample confidence loss. x ij represents the jth real box matched to the ith prior box of signal category m, and the value of 1 indicates matching and the value of 0 indicates non-matching.

[0072] (6) Test: The test set is divided into 7 groups according to the signal-to-noise ratio, which are -9dB, -6dB, -3dB, 0dB, 3dB, 6dB and 9dB. The test set also has 6 radar waveforms, which are consistent with the training set. Each signal has 50 time-frequency graphs under each signal-to-noise ratio, so the test set has a total of 300 time-frequency graphs.

[0073] First, the effective feature layer of the test set is extracted to obtain the adjustment of each prior box and the prediction of all signal categories corresponding to it. Through stacking and activation function, the position information of all prior boxes and the confidence of the predicted category are obtained. Then the format of the predicted box information is converted into the format of the real box information with center point, box width and box height through decoding. Then threshold judgment is performed on all predicted boxes corresponding to each predicted category, and the predicted boxes with confidence greater than the threshold are taken out. The non-maximum suppression algorithm is used to retain the predicted boxes with low overlap and high confidence between the predicted boxes. The retained predicted boxes are sorted in descending order of confidence, and the predicted box information with the highest confidence is taken out, and the previously filled gray bars are removed. Finally, the effective area is identified on the test set time-frequency image, and the predicted box and confidence are drawn.

[0074] (7) Evaluation index: According to the characteristics of image classification, the evaluation index is optimized, and the evaluation mechanism in image classification is adopted. In each time-frequency image in the test set, the prediction box with the highest prediction classification confidence and the corresponding prediction category are selected as the final prediction result of the time-frequency image, as shown in Fig. 2(b). By comparing the prediction category and the true label, the recognition accuracy of a certain waveform under a certain signal-to-noise ratio is the ratio of the number of correctly predicted time-frequency images of the waveform to the total number of time-frequency images of the waveform. The average of the recognition accuracies of each waveform under the same signal-to-noise ratio is the total recognition accuracy under the signal-to-noise ratio.

[0075] (8) Experimental test: Under the condition of 0dB signal-to-noise ratio, the recognition accuracy of BPSK signal is 100%, the recognition accuracy of LFM signal is 100%, the recognition accuracy of P1 signal is 100%, the recognition accuracy of P2 signal is 96%, the recognition accuracy of P3 signal is 100%, and the recognition accuracy of P4 signal is 84%. Under the condition of -3dB signal-to-noise ratio, the recognition accuracy of BPSK signal is 100%, the recognition accuracy of LFM signal is 100%, the recognition accuracy of P1 signal is 54%, the recognition accuracy of P2 signal is 88%, the recognition accuracy of P3 signal is 100%, and the recognition accuracy of P4 signal is 88%.

[0076] To illustrate the effectiveness of the method proposed in the present application, the following comparative experiment is completed: the conventional image classification method uses CNN to recognize the whole time-frequency image, and the data set constructed by the present application is used for training and testing. Finally, compared with the conventional image classification method, the total recognition accuracy of the radar waveform classification method based on effective area recognition is effectively improved, and the experimental results are shown in Table 1:

[0077] Table 1 Total recognition rate comparison

[0078]

[0079] Figure 5 A radar waveform system based on effective area recognition is provided for the embodiment of the present application, as shown in Figure 5 The radar waveform system based on effective area recognition includes a training module 501 and an identification module 502, wherein:

[0080] The training module 501 is used for training the SSD network by taking a plurality of radar emitter time-frequency images as training samples, taking the effective area of a single intra-pulse signal in each radar emitter time-frequency image and the corresponding radar waveform category as labels, and the identification module 502 is used for inputting the radar emitter time-frequency image into the trained preset SSD network to obtain the effective area of the single intra-pulse signal and the corresponding radar waveform category identified by the preset SSD network.

[0081] It can be understood that the radar waveform system based on effective area recognition provided by the present application corresponds to the radar waveform method based on effective area recognition provided by the foregoing embodiments, and the related technical features of the radar waveform system based on effective area recognition can refer to the related technical features of the radar waveform method based on effective area recognition, which will not be repeated here.

[0082] Please refer to Figure 6 , Figure 6 The embodiment of the electronic device provided by the present application is shown in the figure. As shown in Figure 6 , the present application provides an electronic device 600, which comprises a memory 610, a processor 620, and a computer program 611 stored in the memory 610 and executable on the processor 620. When the processor 620 executes the computer program 611, the following steps are implemented: inputting a radar radiation source time-frequency image into a pre-trained SSD network, obtaining an effective area where a single intra-pulse signal is located and a corresponding radar waveform category output by the pre-trained SSD network; the pre-trained SSD network is trained with multiple radar radiation source time-frequency images as training samples, and the effective area where the single intra-pulse signal is located in each radar radiation source time-frequency image and the corresponding radar waveform category are used as labels.

[0083] Please refer to Figure 7 , Figure 7 The embodiment of the computer readable storage medium provided by the present application is shown in the figure. As shown in Figure 7 , the present embodiment provides a computer readable storage medium 700, which stores a computer program 711. When the computer program 711 is executed by a processor, the following steps are implemented: inputting a radar radiation source time-frequency image into a pre-trained SSD network, obtaining an effective area where a single intra-pulse signal is located and a corresponding radar waveform category output by the pre-trained SSD network; the pre-trained SSD network is trained with multiple radar radiation source time-frequency images as training samples, and the effective area where the single intra-pulse signal is located in each radar radiation source time-frequency image and the corresponding radar waveform category are used as labels.

[0084] The radar waveform classification method and system based on effective area recognition provided by the present application have the following advantages:

[0085] (1) For the case that the intra-pulse signal in the radar radiation source time-frequency image is incomplete or contains multiple cycles, the effective area where the single intra-pulse signal is located in the time-frequency image is labeled by a real box, and the network is trained to learn the signal features in this area. When the network is trained, it has the ability to automatically and quickly identify the effective area in the time-frequency image, avoiding the influence of signals in the invalid area on identification, so as to more accurately predict the radar waveform modulation pattern and improve the identification accuracy.

[0086] (2) It is applicable to different size time-frequency images and intrapulse signal distribution in different areas of time-frequency images, and has good adaptability.

[0087] (3) The open system framework using software radio and deep learning can flexibly receive waveforms, update the database in time, has miniaturization characteristics, is rapidly deployed, and can respond to battlefield emergencies at any time.

[0088] It should be noted that in the above embodiments, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0089] Those skilled in the art will understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0090] The present application is described with reference to flowcharts and / or block diagrams of methods, devices (systems) and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing devices to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 means for performing the functions specified in one or more flows and / or blocks.

[0091] These computer program instructions can also be stored in a computer-readable memory that can guide the computer or other programmable data processing devices to work in a specific way, so that the instructions stored in the computer-readable memory produce a product including instruction means, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 means for performing the functions specified in one or more flows and / or blocks.

[0092] These computer program instructions can also be loaded into a computer or other programmable data processing device, so that a series of operation steps are performed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide a means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0093] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0094] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for classifying real-space radar waveforms based on effective region identification. Its features are, include: The radar waveform is received through a software radio platform, and the radar waveform is converted into a time-frequency image using the Choi-Williams distribution. The hardware part of the software radio platform consists of HackRF and a PC. The radar radiation source time-frequency image is input into the trained preset SSD network. The radar radiation source time-frequency image includes cases where the intra-pulse signal is incomplete or contains multiple cycles. The effective region of the single intra-pulse signal output by the preset SSD network and the corresponding radar waveform category are obtained. The preset SSD network is trained using radar waveforms received via a software radio platform and multiple radar source time-frequency images transformed based on the radar waveforms using the Choi-Williams distribution as training samples. The effective region of the signal within a single pulse in each radar source time-frequency image and the corresponding radar waveform category are used as labels. The training samples and labels of the preset SSD network are obtained as follows: Determine multiple signal-to-noise ratios and collect time-frequency images corresponding to multiple radar waveform categories under each signal-to-noise ratio; The effective region containing a single intra-pulse signal in each time-frequency image is marked, as well as the radar waveform category of the single intra-pulse signal; when the intra-pulse signal contained in the time-frequency image has multiple cycles, the effective region containing the most completely displayed intra-pulse signal in the time-frequency image is selected by bounding a box to obtain the effective region target box of the intra-pulse signal, and its radar waveform category is marked; the preset SSD network includes a backbone network, a detection network, and an output network; The backbone network uses a feature pyramid approach to extract multiple effective feature layers from the time-frequency image; The detection network is used to perform two convolutions on each effective feature layer to obtain adjustment information of multiple prior boxes and the radar waveform category corresponding to each prior box; The output network is used to obtain the prediction box with the highest matching degree from multiple prior boxes as the output prediction box; The output network is used to obtain the prediction box with the highest matching degree from multiple prior boxes as the output prediction box, including: For any time-frequency image, calculate the area intersection-union ratio (IOU) between each prior bounding box and the effective target bounding box; Discard prior boxes whose area intersection-union ratio (IOU) with the target boxes of the effective region is less than the IOU threshold, and retain prior boxes whose area intersection-union ratio (IOU) with the target boxes of the effective region is greater than or equal to the IOU threshold; Obtain the radar waveform category corresponding to each retained prior box; The prior bounding box with the highest confidence level for each radar waveform category is obtained as the output predicted bounding box; obtaining the radar waveform category corresponding to each retained prior bounding box includes: For any given prior box corresponding to multiple radar waveform prediction categories, a threshold judgment is performed on all prior boxes corresponding to each prediction category to obtain prior boxes with a confidence level greater than the threshold. The step of obtaining the prior box with the highest confidence level for each radar waveform category as the output prediction box includes: For the multiple retained prior boxes and the radar waveform category corresponding to each prior box, the non-maximum suppression algorithm is used to select the prior box with the highest confidence as the output prediction box; the prediction box with the highest prediction confidence and the corresponding prediction category are selected as the final prediction result of the time-frequency map; by comparing the predicted category with the true label, the recognition accuracy of a certain waveform under each signal-to-noise ratio is the ratio of the number of correctly predicted time-frequency maps of that waveform to the actual number of all time-frequency maps of that waveform; the average of the recognition accuracy of each waveform under the same signal-to-noise ratio is the total recognition accuracy under that signal-to-noise ratio.

2. The actual space radar waveform classification method according to claim 1, characterized in that, Also includes: The predicted bounding box and radar waveform category of the intrapulse signal in each output time-frequency image are compared with the marked effective area and radar waveform category, and the recognition accuracy of the preset SSD network is calculated based on the comparison results. The preset SSD network is repeatedly trained to improve its recognition accuracy until the SSD network converges.

3. A real-space radar waveform classification system based on effective region identification, characterized in that, include: The training module is used to train a preset SSD network using multiple radar radiation source time-frequency images as training samples and the effective area of ​​a single pulse signal in each radar radiation source time-frequency image and the corresponding radar waveform category as labels. The radar radiation source time-frequency images include cases where the pulse signal is incomplete or contains multiple cycles. The preset SSD network is trained by receiving radar waveforms through a software radio platform and using multiple radar radiation source time-frequency images converted based on radar waveforms using the Choi-Williams distribution as training samples. The effective area of ​​the signal within a single pulse in each radar radiation source time-frequency image and the corresponding radar waveform category are used as labels. The training samples and labels of the preset SSD network are obtained as follows: Multiple signal-to-noise ratios (SNRs) are determined, and time-frequency images corresponding to multiple radar waveform categories under each SNR are collected; the effective region containing a single intra-pulse signal in each time-frequency image is labeled, as well as the radar waveform category of the single intra-pulse signal; when the intra-pulse signal contained in the time-frequency image has multiple cycles, the effective region containing the most completely displayed intra-pulse signal in the time-frequency image is selected by bounding a box to obtain the effective region target box of the intra-pulse signal, and its radar waveform category is labeled; the preset SSD network includes a backbone network, a detection network, and an output network; the backbone network extracts multiple effective feature layers of the time-frequency image using a feature pyramid method; the detection network performs two convolutions on each effective feature layer to obtain adjustment information for multiple prior boxes and the radar waveform category corresponding to each prior box; The output network is used to obtain the prediction box with the highest matching degree from multiple prior boxes as the output prediction box; The output network is used to obtain the highest matching prior box from multiple prior boxes as the output predicted box, including: for any time-frequency image, calculating the area intersection-union ratio (IOU) between each prior box and the effective region target box; discarding prior boxes whose IOU with the effective region target box is less than the IOU threshold, and retaining prior boxes whose IOU with the effective region target box is greater than or equal to the IOU threshold; obtaining the radar waveform category corresponding to each retained prior box; obtaining the prior box with the highest confidence in the radar waveform category as the output predicted box; the step of obtaining the radar waveform category corresponding to each retained prior box includes: for multiple radar waveform prediction categories corresponding to any prior box, selecting each prediction category... All prior bounding boxes are thresholded, and prior bounding boxes with a confidence level greater than the threshold are obtained. The step of obtaining the prior bounding box with the highest confidence level for the radar waveform category as the output prediction box includes: for the multiple retained prior bounding boxes and the radar waveform category corresponding to each prior bounding box, using a non-maximum suppression algorithm to select the prior bounding box with the highest confidence level as the output prediction box; selecting the prediction box with the highest prediction confidence level and its corresponding prediction category as the final prediction result for the time-frequency image; by comparing the predicted category with the true label, the recognition accuracy rate for a certain waveform at each signal-to-noise ratio is the ratio of the number of correctly predicted time-frequency images of that waveform to the actual number of all time-frequency images of that waveform; the average of the recognition accuracy rates for each waveform at the same signal-to-noise ratio is the total recognition accuracy rate at that signal-to-noise ratio. The acquisition module is used to input the time-frequency image of the radar radiation source into the trained preset SSD network, and to acquire the effective region of the single pulse signal output by the preset SSD network and the corresponding radar waveform category.

4. A computer-readable storage medium, characterized in that, It stores a computer management program, which, when executed by a processor, implements the steps of the real-space radar waveform classification method based on effective area identification as described in any one of claims 1-2.

Images