Intelligent sea surface ship detection method based on radar RD image under passive jamming condition

Abstract

The invention belongs to the technical field of radar detection, and discloses a passive jamming condition sea surface ship intelligent detection method based on a radar RD image. The method comprises the following steps: step 1, obtaining a sea surface radar RD image data set; dividing a training set and a verification set; 2, constructing a radar RD image ship target lightweight detection model by taking the YOLOv8 target detection model as a basic network; step 3, inputting the training set into the radar RD image ship target lightweight detection model constructed in the step 2 for training to obtain a trained radar RD image ship target lightweight detection model; and step 4, detecting the ship target in the sea surface RD image to be detected by using the trained radar RD image ship target lightweight detection model to obtain the detected ship target. Compared with other existing algorithms, the method has the advantages that the false detection of ship types is obviously reduced, and the method has better detection and recognition capability on various targets, so that the model can more accurately position the targets.

Description

Technical Field

[0001] The invention belongs to the technical field of radar detection, and in particular relates to an intelligent detection method for sea surface ships based on radar RD images under passive interference conditions. Background Art

[0002] Sea surface target detection has important value in many fields such as national defense and military, maritime surveillance, transportation, and resource and environmental protection. Compared with optical detection methods such as visible light imaging, infrared detection, and laser detection, target detection based on radar images has the advantages of all-day, all-weather, long range, and is not restricted by natural conditions such as clouds and fog. Compared with synthetic aperture radar images with large bandwidth, multi-channel and wide-band imaging, radar range-Doppler (RD) images have become the latest development trend in the field of radar image target detection due to their low imaging cost, fast imaging speed, broad application background, and far-reaching practical significance. RD images have extremely low requirements for imaging radars. Radars use microwave pulses to record backscattered echoes from the earth, and obtain radar RD images by performing slow-time pulse compression and fast Fourier transform (FFT) processing on the original echo data. They are the core information source of current marine radars.

[0003] However, the sea surface radar RD image has the characteristics of low resolution, sparse target distribution, small target size, and little feature information. Although the imaging process of radar RD images is simple and real-time, the simplified imaging also leads to low resolution of RD images, blurred edge information of targets and easy to be polluted by background sea clutter. At the same time, the imaging of sea surface radar RD images is highly redundant, the target context information is limited, and the target occupies very few pixels in the image, resulting in limited target detail information. In addition, the rapid development of marine jamming technology in recent years has greatly changed the application scenarios of target detection algorithms. Among them, passive jamming is one of the main challenges facing the current marine target detection of radar RD images. Passive jamming greatly affects the radar system's ability to detect and identify targets by simulating target characteristics, and increases the detection difficulty of target detection algorithms.

[0004] Therefore, ship target detection based on radar RD images is a current hotspot and difficulty. Although there have been some research results, the effectiveness is insufficient. In summary, the proposed method of intelligent detection of sea surface ships based on radar RD images under passive interference conditions has both practical significance and theoretical difficulty. Summary of the invention

[0005] The purpose of the present invention is to provide an intelligent detection method for sea surface ships based on radar RD images under passive interference conditions, so as to solve the problem that the existing radar RD image target detection method has poor detection effect and make the detection result more accurate.

[0006] In order to achieve the above object, the present invention adopts the following technical solution: A method for intelligent detection of sea surface vessels based on radar RD images under passive interference conditions specifically comprises the following steps: Step 1: Generate a range Doppler image using the original radar echo signal, and perform cropping and padding operations on the range Doppler image to obtain a sea surface radar RD image dataset; annotate the targets of the sea surface radar RD image dataset, and divide the training set and the validation set into two sets in proportion; Step 2: Using the YOLOv8 target detection model as the basic network to build a lightweight detection model for ship targets in radar RD images; specifically, it includes the following sub-steps: Step 21: Delete the last layer of standard convolution and C2f modules in the backbone network of the YOLOv8 target detection model, adjust the corresponding part of the Neck feature fusion network, and add a small target detection head in the P2 stage to obtain a preliminary improved network; Step 22: Use the feature transformation module of spatial to depth convolution to replace the standard convolution module of the initial input in the preliminary improved network; Step 23: The C2f module in the P3 and P4 stages of the network obtained in step 22 is improved as follows: specifically, all standard convolutions except the CBS operation and the Spilt operation in the C2f module are replaced with dilated convolutions with a dilation factor of 2; Step 24: embed an efficient channel attention module after the concat operation of P2, P3 and P4 stages in the Neck part of the network obtained in step 23; Step 25: The classification loss function used in the classification loss calculation of the detection head is obtained by weighting the cross entropy loss function; Step 3, input the training set obtained in step 1 into the radar RD image ship target lightweight detection model constructed in step 2 for training, obtain the model parameters that meet the requirements, and verify the effect through the validation set to obtain the trained radar RD image ship target lightweight detection model; Step 4: Use the trained radar RD image ship target lightweight detection model obtained in step 3 to detect the ship target in the sea surface RD image to be detected, and obtain the detected ship target.

[0007] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention provides fine-grained information for the target detection algorithm by compressing the network depth and introducing a feature transformation module of space-to-depth convolution. At the same time, the C2f module is improved by using the dilated convolution to make up for the lack of abstract information caused by the lack of deep features. In addition, by introducing an efficient channel attention module in the feature fusion stage, the model's target recognition and precise positioning capabilities are improved, and the sliding loss function is used to improve the classification loss, thereby improving the algorithm's detection accuracy for difficult samples. Compared with other existing algorithms, the present invention significantly reduces false detections of ship categories and has better detection and recognition capabilities for a variety of targets, enabling the model to more accurately locate targets, and has achieved satisfactory performance in terms of detection accuracy and model lightweighting. DETAILED DESCRIPTION

[0008] The present invention is described in detail below in conjunction with specific implementation modes.

[0009] The passive interference condition provided by the present invention is based on the radar RD image sea surface ship intelligent detection method, which specifically includes the following steps: Step 1: Generate a range Doppler image using the original radar echo signal, and perform cropping and padding operations on the range Doppler image to obtain a sea surface radar RD image dataset; annotate the targets of the sea surface radar RD image dataset, and divide the training set and the validation set into two sets in proportion; Specifically, a total of 4940 images are generated, and the image size is There are more than 9,600 instances in the dataset, including two types of targets: ships and passive interference, and the training set and validation set are divided into 8:2. In this step, the subsequent training difficulties caused by the large aspect ratio of the original image can be solved by cropping and padding operations.

[0010] Step 2: Using the YOLOv8 target detection model as the basic network to build a lightweight detection model for ship targets in radar RD images; specifically, it includes the following sub-steps: Step 21: Delete the last layer of standard convolution and C2f modules in the backbone network of the YOLOv8 target detection model to reduce the number of downsampling times for feature extraction, adjust the corresponding part of the Neck feature fusion network, and add a small target detection head in the P2 stage to obtain a preliminary improved network. The preliminary improved network obtained by this step can retain small target information as much as possible.

[0011] Step 22: Use the feature transformation module of spatial to deep convolution to replace the standard convolution module of the initial input in the preliminary improved network. The process of the feature transformation module of spatial to deep convolution is as follows: Step 221: Segment the original feature map in the input preliminary improved network into a series of sub-feature maps. The specific calculation formula is as follows:

[0012] Where: —Dimensions are The original feature map of is the number of channels of the original feature map; —The width and height of the original feature map; —The scaling factor of downsampling segmentation, in this embodiment, ; —The first Row, No. Sub-feature map of the column, where ; Step 222: Connect all sub-feature maps obtained in step 221 along the channel dimension using the concat operation to obtain a new feature map , new feature map The spatial dimension of (compared to the original feature map) is reduced times, channel dimension magnification times, that is, the new feature map The size is .

[0013] Step 223: New feature map Perform non-strided convolution processing, the number of channels of non-strided convolution is , , and the size is feature map.

[0014] Step 23: The C2f module in the P3 and P4 stages of the network obtained in step 22 is improved as follows: specifically, all standard convolutions except the CBS operation and the Spilt operation in the C2f module are replaced by dilated convolutions with a dilation factor of 2.

[0015] The dilated convolution can expand the receptive field without changing the output resolution and capture contextual information in a larger range.

[0016] Step 24: After the concat operation of the P2, P3 and P4 stages in the Neck part of the network obtained in step 23, an efficient channel attention module is embedded. The operation of the efficient channel attention module is as follows: Step 241: Using two spatial ranges of the pooling kernel , , perform one-to-one one-dimensional feature encoding operations on the input features along the horizontal and vertical coordinates respectively to obtain the intermediate feature map , In this step, long-range dependencies are captured along one spatial direction while precise location information is retained along another spatial direction, which can help the network locate the target more accurately.

[0017] Specifically, the intermediate feature map , In the The output on each channel is as follows (the same is true for other channels):

[0018] Where: -Input feature map in channel Upper height is , width is The characteristic value when ; -Input feature map in channel Upper height is , width is The characteristic value when ; —No. The channels are located at heights of and width is Output when — integers in the variable domain; —The width of the input feature map; —The height of the input feature map.

[0019] Step 242: In order to fully highlight the region of interest and effectively capture the spatial and channel relationships, attention generation is performed on this basis, and the attention generated in step 241 is , The data are input into the one-dimensional depth-wise separable convolution respectively. In addition, since the batch size of the radar RD image target detection algorithm is generally small, batch normalization is used to process the data obtained by the one-dimensional depth-wise separable convolution to obtain the attention weights in the horizontal and vertical directions. Finally, the attention weights in the horizontal and vertical directions are multiplied with the input feature map to obtain a new feature map.

[0020] The specific formula is as follows:

[0021] Where: —Two one-dimensional depthwise separable convolution functions in horizontal and vertical directions; —Intermediate feature map after encoding spatial information horizontally and vertically; —Batch normalization function; —sigmoid function; — horizontal and vertical attention weights; Specifically, the output of the efficient channel attention module on channel C is as follows (the same applies to other channels):

[0022] Where: —The output of the efficient channel attention module on the channel; —Input features of the efficient channel attention module on channel C.

[0023] In this step, the step 241 obtained , Processing sequence information in two directions can achieve more efficient calculations. At the same time, batch normalization can effectively reduce the impact of data distribution differences caused by small batches on network performance and reduce the risk of overfitting.

[0024] Step 25: The classification loss function used in the detection head classification loss calculation is obtained by weighting the cross entropy loss function.

[0025] The calculation formula of the classification loss function is as follows:

[0026] Where: —Classification loss function; — sliding loss function; —Cross entropy loss function.

[0027] The formula of the sliding loss function is as follows:

[0028] Where: —The threshold for judging the difficulty of a sample is equal to the mean of the intersection-over-union ratios of all bounding boxes output by the regression head; —The true category label of the sample after smoothing, the value range is .

[0029] Step 3: Input the training set obtained in step 1 into the radar RD image ship target lightweight detection model constructed in step 2 for training to obtain model parameters that meet the requirements, and verify the effect through the validation set to obtain a trained radar RD image ship target lightweight detection model.

[0030] The network model training and validation set effect evaluation include: the experiment uses the stochastic gradient descent (SGD) method for training, the initial learning rate is set to 0.01, the momentum is set to 0.9, and the warm-up and cosine annealing learning rate optimization strategy is adopted. This strategy makes the learning rate increase linearly from a very small value to the preset value, which can prevent the model from overfitting; the weight decay and initial warm-up momentum of the optimizer are set to 5e-4 and 0.8 respectively, the training batch is set to contain 16 images, and 200 epochs are trained on the training set. To quantify the detection performance of the model, the accuracy ( ), recall rate ( ), Average Precision( ), F1 score and model parameter quantity (Params) are used as evaluation indicators. The detection effect of the method of the present invention on the validation set is shown in Table 1.

[0031] Table 1 Performance comparison of RD image target detection algorithms based on deep learning

[0032] Obviously, the proposed method has a great advantage in F1 score and mAP0.5. Compared with the YOLOv8 prototype, the proposed algorithm performs well in ship category target detection. When the P value decreases by about 2%, the R value increases by 17.9%, the F1score increases by 8.9%, and the mAP0.5 also increases by 4.4%, which greatly improves the model's detection performance for ship targets and reduces the network's false detection of ship targets. The F1 score of all categories increases by 5.1%, and the mAP0.5 increases by 3.4%, proving that the proposed method has good detection and recognition capabilities for a variety of targets on the sea surface RD image dataset, and can locate targets more accurately. At the same time, the model parameters of the proposed method are reduced by 65.1% compared with the target detection algorithm with the smallest number of parameters currently used, making the target detection model more lightweight and easy to embed on mobile devices. In summary, the proposed method improves the model's ability to distinguish between ship targets and passive jammers, while making the position of the target frame more accurate, further improving the model's target detection capability, and is more lightweight than other target detection models.

[0033] Step 4: Use the trained radar RD image ship target lightweight detection model obtained in step 3 to detect the ship target in the sea surface RD image to be detected, and obtain the detected ship target.

[0034] In summary, the present invention provides fine-grained information for the target detection algorithm by compressing the network depth and introducing a feature transformation module of space-to-depth convolution. At the same time, the C2f module is improved by using the hole convolution to make up for the lack of abstract information caused by the lack of deep features. In addition, by introducing an efficient channel attention module in the feature fusion stage, the model's target recognition and precise positioning capabilities are improved, and the classification loss is improved by using the sliding loss function, which improves the algorithm's detection accuracy for difficult samples. Compared with other existing algorithms, the present invention significantly reduces false detections of ship categories, and has better detection and recognition capabilities for a variety of targets, enabling the model to more accurately locate targets, and has achieved satisfactory performance in terms of detection accuracy and model parameter quantity.

Claims

1. An intelligent detection method for sea surface vessels based on radar RD images under passive interference conditions, characterized in that: The specific steps include: Step 1: Generate a range Doppler image using the original radar echo signal, and perform cropping and padding operations on the range Doppler image to obtain a sea surface radar RD image dataset; annotate the targets of the sea surface radar RD image dataset, and divide the training set and the validation set into two sets in proportion; Step 2: Using the YOLOv8 target detection model as the basic network to build a lightweight detection model for ship targets in radar RD images; specifically, it includes the following sub-steps: Step 21: Delete the last layer of standard convolution and C2f modules in the backbone network of the YOLOv8 target detection model, adjust the corresponding part of the Neck feature fusion network, and add a small target detection head in the P2 stage to obtain a preliminary improved network; Step 22: Use the feature transformation module of spatial to depth convolution to replace the standard convolution module of the initial input in the preliminary improved network; Step 23: The C2f module in the P3 and P4 stages of the network obtained in step 22 is improved as follows: specifically, all standard convolutions except the CBS operation and the Spilt operation in the C2f module are replaced with dilated convolutions with a dilation factor of 2; Step 24: embed an efficient channel attention module after the concat operation of the P2 stage, P3 stage, and P4 stage in the Neck part of the network obtained in step 23; Step 25: The classification loss function used in the classification loss calculation of the detection head is obtained by weighting the cross entropy loss function; Step 3, input the training set obtained in step 1 into the radar RD image ship target lightweight detection model constructed in step 2 for training, obtain the model parameters that meet the requirements, and verify the effect through the validation set to obtain the trained radar RD image ship target lightweight detection model; Step 4: Use the trained radar RD image ship target lightweight detection model obtained in step 3 to detect the ship target in the sea surface RD image to be detected, and obtain the detected ship target.

2. The method for intelligent detection of sea surface vessels based on radar RD images under passive interference conditions as claimed in claim 1, characterized in that: In step 22, the process of the feature transformation module of the space-to-depth convolution is as follows: Step 221: dividing the original feature map in the input preliminary improvement network into a series of sub-feature maps; Step 222: Connect all sub-feature maps obtained in step 221 along the channel dimension using the concat operation to obtain a new feature map ; Step 223: New feature map Perform non-strided convolution.

3. The method for intelligent detection of sea surface vessels based on radar RD images under passive interference conditions as claimed in claim 2, characterized in that: In step 24, the operation of the efficient channel attention module is as follows: Step 241: Using two spatial ranges of the pooling kernel , , perform one-to-one one-dimensional feature encoding operations on the input features along the horizontal and vertical coordinates respectively to obtain the intermediate feature map , ;in, , are the width and height of the input feature map respectively; Step 242: Use the following two formulas to , The attention weights in the horizontal and vertical directions are obtained respectively; Where: —Two one-dimensional depthwise separable convolution functions in horizontal and vertical directions; —Intermediate feature map after encoding spatial information horizontally and vertically; —Batch normalization function; —sigmoid function; — horizontal and vertical attention weights; The horizontal and vertical attention weights are multiplied with the input feature map to obtain a new feature map. The output of the efficient channel attention module on channel C is expressed as follows: Where: —The output of the efficient channel attention module on the channel; —Input features of the efficient channel attention module on channel C.

4. The method for intelligent detection of sea surface vessels based on radar RD images under passive interference conditions as claimed in claim 2, characterized in that: In step 25, the calculation formula of the classification loss function is as follows: Where: —Classification loss function; — sliding loss function; —Cross entropy loss function; The formula of the sliding loss function is as follows: Where: —Threshold for judging the difficulty of samples; —The true category label of the sample after smoothing, the value is .

Images