Intelligent detection method of lightweight sea surface vessels based on radar RD images

By improving the YOLOv8 target detection model, a lightweight detection model for radar RD image ship targets was constructed, which solved the problems of poor detection effect and limited detection speed in sea surface scenes, and achieved more accurate and efficient sea surface ship detection .

CN118762027BActive Publication Date: 2024-11-29ROCKET FORCE UNIV OF ENG
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202411252614.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-09
Publication Date
2024-11-29
Estimated Expiration
2044-09-09

AI Technical Summary

Technical Problem

The existing radar RD image target detection method has poor detection effect in sea surface scenes and has limited detection speed, making it difficult to accurately detect sea surface ships in interference or high-intensity sea clutter scenarios.

Method used

Based on the YOLOv8 target detection model, a lightweight detection model for radar RD image ship targets was constructed. By expanding the number of network channels, reducing the number of downsamplings, using combined cavity convolution and global context modules, the lightweight combined cavity module RCDGhost is designed. , and optimize model parameter quantity and calculation cost through structural reparameterization and query head design.

Benefits of technology

More accurate sea surface vessel detection under disturbing conditions is achieved, which reduces computational redundancy, improves detection speed, and significantly reduces the number of model parameters and floating-point operations while maintaining model accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure SMS_18
    Figure SMS_18
  • Figure SMS_33
    Figure SMS_33
  • Figure SMS_36
    Figure SMS_36
Patent Text Reader

Abstract

The present invention discloses a lightweight sea surface ship intelligent detection method based on radar RD images: step 1, using the original radar echo signal to generate a range Doppler image to obtain a sea surface radar RD image data set; marking the targets of the sea surface radar RD image data set, and dividing the training set and the verification set; step 2, building a radar RD image ship target lightweight detection model based on the YOLOv8 target detection model; step 3, inputting the training set into the radar RD image ship target lightweight detection model for training; step 4, using the trained radar RD image ship target lightweight detection model to detect the ship target in the sea surface RD image to be detected, and obtaining the ship target. The present invention reduces the target information loss while reducing the network parameters, and compensates for the accuracy reduction caused by lightweighting while being friendly to the parameter amount and the number of operations.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of radar detection, and in particular relates to an intelligent detection method for lightweight sea surface vessels based on radar RD images. 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.

[0003] Traditional radar RD image detection methods are based on constant false alarm rate (CFAR) algorithms, which use constant false alarm rate judgment thresholds for detection. However, this type of algorithm based on a single pixel has large detection redundancy and will cause a large number of false detections in interference or high-intensity sea clutter scenes. In recent years, inspired by the achievements of computer vision, RD image intelligent detection algorithms based on deep learning have opened up broader prospects for this direction. However, it is often difficult to achieve good results by directly applying existing general models to sea surface RD image target detection scenarios. Sea surface RD images have the characteristics of low resolution, little feature information, sparse target distribution, and small target size. Considering its image characteristics, it is urgent to design a more accurate and faster RD image target detection method for this scene. Summary of the invention

[0004] The purpose of the present invention is to provide a lightweight sea surface vessel intelligent detection method based on radar RD images to solve the problems of poor detection effect and limited detection speed of existing radar RD image target detection methods, so as to make the detection results more accurate and reduce calculation redundancy.

[0005] In order to achieve the above object, the present invention adopts the following technical solution:

[0006] A lightweight sea surface ship intelligent detection method based on radar RD images specifically comprises the following steps:

[0007] 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;

[0008] Step 2: Using the YOLOv8 target detection model as the basic network, a lightweight detection model for ship targets in radar RD images is constructed;

[0009] 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 model parameters that meet the requirements, and verify the effect through the verification set;

[0010] 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.

[0011] Furthermore, step 2 specifically includes the following sub-steps:

[0012] Step 21: Keep the P3 and P4 stages of the YOLOv8 target detection model, expand the neck network and detection head of the P2 stage; use the padding operation in the standard convolution of the P4 stage, and delete the P5 stage in the backbone network to obtain a preliminary improved network;

[0013] Step 22: Adjust the number of model channels of the preliminary improved network obtained in step 21, and double the number of output channels of each operation;

[0014] Step 23: Replace the C2f modules in the P2 and P3 stages of the network for adjusting the number of channels obtained in step 22 with lightweight combined hole modules to obtain an improved network;

[0015] Step 24: Set the padding parameters of the two parallel decoupling heads in the improved network obtained in step 23 to 1, and embed a query head in parallel with the classification head and the regression head in the P3 stage and P4 stage of the improved network respectively. The input of the query head is the feature map output by the P3 stage or P4 stage of the backbone network of the improved network; set the convolution kernel parameter of the query head to 3, the step size to 1, and the padding parameter to 1, and obtain the lightweight detection model of ship targets in radar RD images.

[0016] Furthermore, the specific operations of step 23 are as follows:

[0017] Step 231: Convert the standard convolution in the P2 stage and the C2f module in the P3 stage of the backbone network of the network with adjusted channel number obtained in step 22 into a combined dilated convolution to obtain a network with improved convolution. The conversion formula from the standard convolution to the combined dilated convolution is as follows:

[0018]

[0019] Where:

[0020] - Input is The weight is , standard convolution with a dilation rate of 1;

[0021] —Combination function, obtained by connecting an average pooling layer with a convolution kernel size of 5*5 to a 1*1 convolution layer;

[0022] x—input;

[0023] —Trainable weights for combining dilated convolutions;

[0024] - Input is The weight is , standard convolution with a dilation rate of 3;

[0025] Step 232: In the convolution-improved network obtained in step 231, insert the combined hole convolution before and after the global context module, which is specifically connected to the 1*1 convolution layer through global average pooling, and then added to the shortcut back to the main feature stream;

[0026] Step 233: Perform a structural reparameterization operation on the Bottleneck module in the network obtained in step 232 to obtain an improved network.

[0027] Furthermore, in step 3, the training process includes the following:

[0028] (1) Based on the anchor-free detection head of the improved network obtained in step 2, the minimum regression range in the P3 stage is set to , then let the target in stage P3 be The true bounding box of , from the feature map to the center of the target truth box in the P3 stage The minimum distance mapping is , each feature position and The minimum distance between the mapping values as follows:

[0029]

[0030] Where:

[0031] — is the characteristic position and The mapping value of the minimum distance between them;

[0032] -Target The coordinate value of the center of the true bounding box;

[0033] -Target The width and height of the ground-truth bounding box;

[0034] (2) Make the distance smaller than the minimum regression range The position is 1, and the rest are 0. According to the mapping value of the minimum distance Encode the query mapping value of the real value and then get the target The target sparse feature map of the query head ,in, is the set of all real numbers, is the height of the target sparse feature map, is the width of the target sparse feature map; the query mapping value of the true value is specifically:

[0035]

[0036] Where: —The query mapping value of the true value, which is used as the feature position in the target sparse feature map The value of

[0037] (3) Perform the same operations (1) and (2) on the P4 stage as for the P3 stage;

[0038] (4) The AdamW optimizer is used for training. The overall loss function during the training process is as follows, and the loss function Focalloss is used to train the query head;

[0039]

[0040] Where:

[0041] —respectively, the output of classification prediction, the output of bounding box regression prediction, and the output of query prediction;

[0042] — respectively The corresponding truth value mapping;

[0043] — classification loss;

[0044] —Bounding box regression loss;

[0045] - Check the loss;

[0046] —are the gains of classification loss, bounding box regression loss, and query loss, respectively.

[0047] Furthermore, step 4 specifically includes the following operations:

[0048] (1) Select the prediction score greater than the threshold The location of P3 is used as the query area, and the query of the P3 stage target Mapped to the four nearest neighbors in the P2 stage as key positions :

[0049]

[0050] in, —The grid positions of the four nearest neighbors in the P2 phase;

[0051] (2) All the Records are key-value position sets , then the classification head and regression head in the P2 stage only query the key value position set The corresponding area is used for target detection;

[0052] (3) Perform the same operations (1) and (2) as for the P3 stage for the P4 stage.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] 1. Aiming at the detection difficulties caused by the unique characteristics of RD image target detection under interference conditions, the present invention constructs a lightweight detection model for radar RD image ship targets based on the YOLOv8 architecture, expands the number of network channels, and reduces the number of downsampling and convolution padding operations to reduce the loss of target information while reducing network parameters.

[0055] 2. In view of the increase in computational complexity caused by expanding the network feature graph and to adapt to efficient deployment on hardware devices in the future, the present invention designs a lightweight combined hole module RCDGhost, which realizes feature reuse through re-parameterization, reduces the number of network model parameters and the number of floating-point operations (GFLOPs), and uses combined hole convolution to integrate deep information, thereby compensating for the loss of accuracy caused by lightweighting while being friendly to the number of parameters and the number of operations.

[0056] 3. In view of the decrease in detection network reasoning speed caused by the expansion of feature maps, the present invention utilizes the sparse distribution of targets in RD images and uses deep features to guide the calculation of shallow features, thereby solving the problem of redundant calculation of high-resolution features and accelerating the model reasoning speed. DETAILED DESCRIPTION

[0057] The present invention is described in detail below in conjunction with specific implementation modes.

[0058] The lightweight sea surface ship intelligent detection method based on radar RD image provided by the present invention specifically comprises the following steps:

[0059] 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;

[0060] The cropping and padding operations in this step can solve the training difficulties caused by the large aspect ratio of the original image.

[0061] Step 2: Use the YOLOv8 target detection model as the basic network to build a lightweight detection model for radar RD image ship targets.

[0062] In this step, the YOLOv8 target detection model is used as the basic network of the lightweight detection model for ship targets in radar RD images. ‌‌YOLOv8 is a deep learning model for target detection, instance segmentation and classification. It is the latest development of the YOLO series of models. The YOLOv8 target detection model includes a backbone network, a neck and a detection head. In the target detection of the YOLOv8 target detection model, the backbone network is used to extract features from the input image and output feature maps of different scales and resolutions. The specific operation is to use a powerful neural network module to perform convolution operations on the input image to continuously obtain feature maps. The neck is used to perform feature fusion on the feature maps of different scales and resolutions output by the backbone network. The specific operation is to stack the feature maps by upsampling and downsampling the feature maps, thereby obtaining rich semantic information, and finally output three layers of feature maps of different scales. The detection head is used to complete the classification and regression tasks of the three layers of feature maps of different scales output by the neck in a decoupled manner. The specific operation is to use the convolution operation and the DFL module to decode the classification information and regression information of the three layers of feature maps of different scales, and output the candidate box and classification confidence information.

[0063] Specifically, step 2 includes the following sub-steps:

[0064] Step 21: Keep the P3 and P4 stages of the YOLOv8 target detection model, expand the neck network and detection head of the P2 stage to improve the feature map resolution; use the padding operation in the standard convolution of the P4 stage, which can keep the spatial resolution of the feature map unchanged and reduce the time and memory consumed by the feature map expansion. At the same time, delete the P5 stage in the backbone network to reduce the number of model downsampling; get a preliminary improved network;

[0065] Step 22: Adjust the number of model channels for the preliminary improved network obtained in step 21, and double the number of output channels for each step. This step can make up for the information loss caused by the reduction in the number of downsampling times, and obtain a network with adjusted channel numbers;

[0066] Step 23: replacing the C2f modules in the P2 and P3 stages of the network for adjusting the number of channels obtained in step 22 with the lightweight combined void module (RCDGhost) designed by the present invention, to obtain an improved network;

[0067] The specific operations of step 23 are as follows:

[0068] Step 231: convert the standard convolution in the C2f module in the P2 stage and the P3 stage in the backbone network of the network with adjusted channel number obtained in step 22 into combined hole convolution to obtain a network with improved convolution.

[0069] Specifically, the formula for converting standard convolution to combined dilated convolution is as follows:

[0070]

[0071] Where:

[0072] - Input is The weight is , standard convolution with a dilation rate of 1;

[0073] —Combination function, obtained by connecting an average pooling layer with a convolution kernel size of 5*5 to a 1*1 convolution layer;

[0074] x—input;

[0075] —Trainable weights for combining dilated convolutions;

[0076] - Input is The weight is , standard convolution with a dilation rate of 3;

[0077] The combined atrous convolution can combine the two convolution results with atrous rates of 1 and 3 through weighted fusion, so that the network can utilize information in receptive fields of different scales.

[0078] Step 232: In the convolution-improved network obtained in step 231, insert the combined dilated convolution before and after the global context module. The global context module is specifically connected to the 1*1 convolution layer (used to compress the input features) through global average pooling, and then added to the shortcut back to the main feature stream. When the global information obtained in this step is available, the combined function Able to make more stable predictions.

[0079] Step 233: Perform a structural reparameterization operation on the Bottleneck module in the network obtained in step 232 to obtain an improved network. The introduction of the structural reparameterization operation in this step can optimize the bottleneck module to achieve implicit feature reuse, thereby further saving computational costs.

[0080] Preferably, the structural reparameterization operation is different in the subsequent training process and the detection process: in the training process, the input features of the Bottleneck module are processed using deep convolution and batch normalization respectively and activated after Add to obtain the actual input features; in the detection process, the input features of the Bottleneck module are processed using deep convolution and batch normalization and activated to obtain the actual input features.

[0081] Specifically, the structural reparameterization operation during training is expressed as:

[0082]

[0083] Where:

[0084] y — output of the structure reparameterization operation;

[0085] Add—add element by element;

[0086] x — input features of the Bottleneck module;

[0087] —Deep convolutional module;

[0088] —Batch Normalization module;

[0089] —Structural reparameterization operation.

[0090] From the above, the operation of step 23 achieves the reduction of network model parameters and floating point operations (GFLOPs) while maintaining the model accuracy.

[0091] Step 24: Set the padding parameters of the two parallel decoupling heads in the improved network obtained in step 23 to 1, and embed a query head in parallel with the classification head and regression head (i.e., two parallel decoupling heads) in the P3 and P4 stages of the improved network, respectively. The input of the query head is the feature map output by the P3 or P4 stage of the backbone network of the improved network. Set the convolution kernel parameter of the query head to 3, the step size to 1, and the padding parameter to 1, and obtain the lightweight detection model of ship targets in radar RD images.

[0092] 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 verification set;

[0093] The training process includes the following:

[0094] (1) Based on the anchor-free detection head of the improved network obtained in step 2, the minimum regression range in the P3 stage is set to , then let the target in stage P3 be The true bounding box of , from the feature map to the center of the target truth box in the P3 stage The minimum distance mapping is , each feature position and The minimum distance between the mapping values as follows:

[0095]

[0096] Where:

[0097] — is the characteristic position and The mapping value of the minimum distance between them;

[0098] -Target The coordinate value of the center of the true bounding box;

[0099] -Target The width and height of the ground-truth bounding box;

[0100] (2) Make the distance smaller than the minimum regression range The position is 1, and the rest are 0. According to the mapping value of the minimum distance Encode the query mapping value of the real value and then get the target The target sparse feature map of the query head , is the set of all real numbers, is the height of the target sparse feature map, is the width of the target sparse feature map; the query mapping value of the true value is specifically:

[0101]

[0102] Where: —The query mapping value of the true value, which is used as the feature position in the target sparse feature map The value of

[0103] (3) Perform the same operations (1) and (2) as for the P3 stage for the P4 stage.

[0104] (4) The AdamW optimizer is used for training, the initial learning rate is set to 0.01, the momentum is set to 0.9, the warm-up and cosine annealing learning rate optimization strategy is adopted, the training batch is set to contain 16 images, and 150 epochs are trained on the training set. The overall loss function during the training process is as follows, and the loss function Focalloss is used to train the query head;

[0105]

[0106] Where:

[0107] —respectively, the output of classification prediction, the output of bounding box regression prediction, and the output of query prediction;

[0108] — respectively The corresponding truth value mapping;

[0109] — classification loss;

[0110] —Bounding box regression loss;

[0111] - Check the loss;

[0112] — are the gains of classification loss, bounding box regression loss, and query loss. is 0.5, is 7.5, is 1.2. They are used to balance the losses of different tasks in the decoupling head.

[0113] Preferably, the classification loss uses the cross entropy loss VFL (Varifocal Loss), and the bounding box regression loss uses the combined loss of CIOU (Complete IOU) and DFL (Distributed Focal).

[0114] 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.

[0115] Step 4 specifically includes the following operations:

[0116] (1) Select the prediction score greater than the threshold The location of P3 is used as the query area, and the query of the P3 stage target Mapped to the four nearest neighbors in the P2 stage as key positions :

[0117]

[0118] in, —Grid positions of the four nearest neighbors in the P2 phase.

[0119] (2) All the Records are key-value position sets , then the classification head and regression head in the P2 stage only query the key value position set The corresponding area is used for target detection.

[0120] (3) Perform the same operations (1) and (2) as for the P3 stage for the P4 stage.

[0121] The following compares the method of the present invention with some currently proposed RD image target detection algorithms based on deep learning, including YOLOv3, YOLOv3-Tiny, YOLOv5 and YOLOv8 prototypes, in terms of model accuracy and lightweightness, aiming to verify that the proposed model has further achieved lightweight while improving accuracy compared to other models, has lower computing resource requirements for the hardware platform, and is more suitable for sea surface target detection tasks in actual scenarios.

[0122] Table 1 Performance comparison of lightweight ship target detection model in radar RD image on dataset 1

[0123]

[0124] Table 1 shows the mAP0.5, mAP0.5:0.95, model parameter quantity, number of floating-point operations, and number of frames processed per second of different algorithm models on the sea surface RD image dataset constructed by the present invention as the final detection comparison indicators. In order to simplify the experimental indicators, mAP0.5 and mAP0.5:0.95 are used for accuracy evaluation to reflect the target detection and positioning capabilities of the algorithm.

[0125] Obviously, the lightweight detection model of ship targets in radar RD images constructed in the present invention realizes the lightweight of the model on the basis of the YOLOv8 prototype. Compared with the YOLOv8 prototype, the proposed algorithm can basically maintain the target detection performance and even improve the target detection accuracy. Its full-category mAP0.5 increased by 1.1%, and mAP0.5:0.95 increased by 5.1%. The query box design greatly improves the accuracy of target positioning. At the same time, YOLO-Q greatly reduces the parameter amount and computational cost of the detection model and improves the model's rapidity. As shown in the table, its parameter amount decreased by 78.3%, and FLOPs decreased by 2.3G, which is the best in the existing YOLO algorithm.

Claims

1. A lightweight sea surface ship intelligent detection method based on radar RD images, 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: Keep the P3 and P4 stages of the YOLOv8 target detection model, expand the neck network and detection head of the P2 stage; use the padding operation in the standard convolution of the P4 stage, and delete the P5 stage in the backbone network to obtain a preliminary improved network; Step 22: Adjust the number of model channels of the preliminary improved network obtained in step 21, and double the number of output channels of each operation; Step 23: Replace the C2f modules in the P2 and P3 stages of the network with adjusted channel numbers obtained in step 22 with lightweight combined hole modules to obtain an improved network; the specific operations are as follows: Step 231: Convert the standard convolution in the P2 stage and the C2f module in the P3 stage of the backbone network of the network with adjusted channel number obtained in step 22 into a combined dilated convolution to obtain a network with improved convolution. The conversion formula from the standard convolution to the combined dilated convolution is as follows: Where: - Input is The weight is , standard convolution with a dilation rate of 1; —Combination function, obtained by connecting an average pooling layer with a convolution kernel size of 5*5 to a 1*1 convolution layer; x—input; —Trainable weights for combining dilated convolutions; - Input is The weight is , standard convolution with a dilation rate of 3; Step 232: In the convolution-improved network obtained in step 231, insert the combined hole convolution before and after the global context module, which is specifically connected to the 1*1 convolution layer through global average pooling, and then added to the shortcut back to the main feature stream; Step 233: performing a structural reparameterization operation on the Bottleneck module in the network obtained in step 232 to obtain an improved network; Step 24: Set the padding parameters of the two parallel decoupling heads in the improved network obtained in step 23 to 1, and embed a query head in parallel with the classification head and the regression head in the P3 stage and the P4 stage of the improved network respectively, and the input of the query head is the feature map output by the P3 stage or the P4 stage of the backbone network of the improved network; The convolution kernel parameter of the query head is set to 3, the step size is 1, and the padding parameter is set to 1, and the lightweight detection model of ship targets in radar RD images is obtained; 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 model parameters that meet the requirements, and verify the effect through the verification set; 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 lightweight sea surface vessels based on radar RD images as claimed in claim 1, characterized in that: In step 3, the training process includes the following: (1) Based on the anchor-free detection head of the improved network obtained in step 2, the minimum regression range in the P3 stage is set to , then let the target in stage P3 be The true bounding box of , from the feature map to the center of the target truth box in the P3 stage The minimum distance mapping is , each feature position and The minimum distance between the mapping values as follows: ; Where: — is the characteristic position and The mapping value of the minimum distance between them; -Target The coordinate value of the center of the true bounding box; -Target The width and height of the ground-truth bounding box; (2) Make the distance smaller than the minimum regression range The position is 1, and the rest are 0. According to the mapping value of the minimum distance Encode the query mapping value of the real value and then get the target The target sparse feature map of the query head ,in, is the set of all real numbers, is the height of the target sparse feature map, is the width of the target sparse feature map; the query mapping value of the true value is specifically: Where: —The query mapping value of the true value, which is used as the feature position in the target sparse feature map The value of (3) Perform the same operations (1) and (2) on the P4 stage as for the P3 stage; (4) The AdamW optimizer is used for training. The overall loss function during the training process is as follows, and the loss function Focalloss is used to train the query head; Where: —respectively, the output of classification prediction, the output of bounding box regression prediction, and the output of query prediction; — respectively The corresponding truth value mapping; — classification loss; —Bounding box regression loss; - Check the loss; —are the gains of classification loss, bounding box regression loss, and query loss, respectively.

3. The method for intelligent detection of lightweight sea surface vessels based on radar RD images as claimed in claim 2, characterized in that: Step 4 specifically includes the following operations: (1) Select the prediction score greater than the threshold The location of P3 is used as the query area, and the query of the P3 stage target Mapped to the four nearest neighbors in the P2 stage as key positions : in, —The grid positions of the four nearest neighbors in the P2 phase; (2) All the Records are key-value position sets , then the classification head and regression head in the P2 stage only query the key value position set The corresponding area is used for target detection; (3) Perform the same operations (1) and (2) as for the P3 stage for the P4 stage.

Citation Information

Patent Citations

  • Millimeter wave radar target detection method based on lightweight network

    CN118425911A