Radar image target detection method based on improved YOLOv8 network model

Abstract

The invention discloses a radar image target detection method based on an improved YOLOv8 network model, and the method comprises the following steps: S1, collecting radar image samples, carrying out the preprocessing and enhancement of the collected radar image samples, and putting the radar image samples into a sample data set; s2, marking a target object in a radar image sample in the sample data set, and dividing the marked sample data set into a training set and a test set; s3, training the improved YOLOv8 network model based on the training sample, and testing the trained improved YOLOv8 network model based on the test sample until the recognition precision of the improved YOLOv8 network model reaches the required recognition precision; and S4, inputting the acquired radar image into the trained improved YOLOv8 network model, and outputting a target object in the radar image. The target detection method has higher resolution and can provide enough details and accuracy; and target detection is carried out in a strong interference environment, so that the detection precision is ensured.

Description

Technical Field

[0001] The present invention belongs to the field of radar target detection, in particular to a radar image target detection method based on an improved YOLOv8 network model. Background Art

[0002] In the field of national defense and security, the monitoring and control of small and medium-sized drones is crucial, especially when these drones approach or may cross the restricted area line. In order to effectively respond to this challenge, it is necessary to identify drones and restricted areas based on radar data, and then track the identified drones.

[0003] Traditional radar target recognition methods are mainly based on statistical characteristics. They process the received radar echo signals forward, optimize the signal quality, and then apply threshold detection algorithms, such as the constant false alarm rate detector (CFAR), to detect whether there is a target in the echo. This radar target detection method has significant advantages in stability and adaptability and is suitable for a variety of conventional application scenarios.

[0004] However, the resolution of the above radar target recognition methods is generally low and cannot provide sufficient details and accuracy; in some complex target scenes or applications that require detailed classification, the target recognition capability advantage will be limited; because its target detection relies on the distinction between signal strength and background noise, it performs poorly in strong interference environments. Summary of the invention

[0005] The present invention provides a radar image target detection method based on an improved YOLOv8 network model, aiming to improve the above problems.

[0006] The present invention is implemented as follows: a radar image target detection method based on an improved YOLOv8 algorithm, the method comprising the following steps:

[0007] S1, collect radar image samples, preprocess and enhance the collected radar image samples, and put them into a sample data set;

[0008] S2, labeling the target objects in the radar image samples in the sample data set, and dividing the labeled sample data set into a training set and a test set;

[0009] S3, training the improved YOLOv8 network model based on the training samples, and testing the trained improved YOLOv8 network model based on the test samples, until the recognition accuracy of the improved YOLOv8 network model reaches the required recognition accuracy;

[0010] S4. Input the collected radar image into the trained improved YOLOv8 network model and output the target object in the radar image.

[0011] Furthermore, the radar image samples include restricted areas for large targets, medium-sized UAVs for medium-sized targets, and small UAVs for small targets.

[0012] Furthermore, the improved YOLOv8 network model includes:

[0013] It consists of backbone network, neck part and head part;

[0014] Among them, the backbone network consists of a sequentially connected convolutional layer Conv, a dynamic convolution ODConv2d, a convolutional layer c2f, and an SPPF module.

[0015] Furthermore, the Head part consists of four detection heads, each of which consists of a convolutional layer Conv and a convolutional layer Con2v.

[0016] Furthermore, the loss value of the improved YOLOv8 network model is L = γ·L SIoU , where L SIoU is the SIOU loss value, and γ is the adjustment coefficient.

[0017] Furthermore, the adjustment coefficient γ=(Cw·h), where C is a constant slightly larger than the maximum area of ​​the set prediction detection frame, w is the width of the prediction detection frame where the target object is located, and h is the height of the prediction detection frame where the target object is located.

[0018] Furthermore, the SIOU loss value L SIoU The calculation formula is as follows:

[0019]

[0020] Among them, Ω is the shape loss, that is, the difference between the length and width ratios of the real detection box and the predicted detection box, Δ is the distance loss, that is, the difference between the minimum circumscribed rectangular area of ​​the real detection box and the predicted detection box; L IoU is the IoU loss.

[0021] Furthermore, the preprocessing process of radar image samples is as follows:

[0022] The collected radar image samples are first converted into grayscale images using the OpenCV library, and then the radar image is subjected to local histogram equalization and then converted back to RGB images.

[0023] Furthermore, the radar image samples are enhanced by using at least one of the following methods:

[0024] (a) Rotate the radar image based on the set rotation angle;

[0025] (b) Flipping the radar image based on the set up, down, left, and right flip probabilities;

[0026] (c) performing a radial transformation on the radar image based on a set affine transformation probability;

[0027] (d) performing hybrid overlap on the radar images based on a set hybrid overlap probability;

[0028] (e) Random mosaics of radar images.

[0029] The radar image target detection method based on the improved YOLOv8 algorithm proposed in the present invention has higher resolution and can provide sufficient details and accuracy; moreover, the method can perform target detection in a strong interference environment to ensure detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 A flow chart of a radar image target detection method based on an improved YOLOv8 network model provided by an embodiment of the present invention;

[0031] Figure 2 A schematic diagram of the structure of an improved YOLOv8 network model provided by an embodiment of the present invention;

[0032] Figure 3 A schematic diagram of image data after sample data enhancement provided by an embodiment of the present invention;

[0033] Figure 4 A histogram comparison diagram of an image before and after data enhancement provided by an embodiment of the present invention;

[0034] Figure 5 A schematic diagram of gradually multiplying the convolution kernel by the attention of ODConv2d provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0035] The specific implementation modes of the present invention are further explained in detail below by describing the embodiments with reference to the accompanying drawings, so as to help those skilled in the art to have a more complete, accurate and in-depth understanding of the inventive concept and technical solution of the present invention.

[0036] Figure 1 A flow chart of a radar image target detection method based on an improved YOLOv8 network model provided in an embodiment of the present invention, the method comprising the following steps:

[0037] S1. Collect radar image samples, including: large target restricted area, medium target medium-sized UAV and small target small UAV, pre-process and enhance the collected radar image samples, put them into the sample data set, increase the sample data scale and improve the sample data quality;

[0038] In the embodiments of the present invention, the data enhancement methods mainly include the following:

[0039] (1) Adjust the brightness of the radar image and stabilize the light source.

[0040] (2) Set the radar image scaling to take into account small target objects.

[0041] (3) Setting the rotation angle of the radar image, the probability of image flipping up and down, left and right, the probability of image affine transformation, the probability of image mixing and overlapping, image mosaic, etc., to expand the image data set while retaining the image features to the greatest extent, such as Figure 3 shown.

[0042] Specifically, the collected radar image is first converted into a grayscale image using the OpenCV library, and then the radar image is subjected to local histogram equalization and then converted back to an RGB image. Figure 4 shown.

[0043] S2. Label the samples in the sample data set, and divide the labeled sample data set into a training set, a validation set, and a test set.

[0044] Each radar image contains small, medium, and large targets. Category label 0 identifies small targets, such as small drones, with pixel values ​​between (0-10)*(10-20); category label 1 identifies medium targets, such as medium drones, with pixel values ​​between (20-30)*(100-120); category label 3 identifies large targets, such as restricted area lines, with pixel values ​​between (20-30)*(580-640). The sizes of different targets vary greatly.

[0045] After the sample description is completed, experienced radar signal processing engineers use Labelme to annotate the targets according to the radar images, and divide the data set into training set, test set, and validation set in a ratio of 8:1:1. The annotated data is stored in the corresponding folder in txt format. The above steps constitute an image data set that can be used by the YOLOv8 model.

[0046] S3, training the improved YOLOv8 network model based on the training samples, and testing the trained improved YOLOv8 network model based on the test samples, until the recognition accuracy of the improved YOLOv8 network model reaches the required recognition accuracy, and the training of the improved YOLOv8 network model is completed;

[0047] S4. Input the collected radar image into the trained improved YOLOv8 network model and output the target in the radar image.

[0048] Establish an improved YOLOv8 network model, initialize the network model, import the training set into the model, set hyperparameters, and train the model. The YOLOv8 network model consists of: backbone network, neck part, and head part, such as Figure 2 As shown in the figure, the backbone network consists of a sequentially connected convolution layer Conv, a dynamic convolution ODConv2d, a convolution layer c2f, and an SPPF module. The convolution layer of the second layer of the backbone network is replaced with ODConv2d dynamic convolution, and a multi-dimensional attention mechanism and parallel strategy are introduced to improve the detection accuracy and generalization ability of the model.

[0049] Specifically, ODConv2d is a more general and elegant dynamic convolution design that uses a multi-dimensional attention mechanism and a parallel strategy to learn the complementary attention of convolution kernels along four dimensions in any convolution layer. Specifically, dynamic convolution ODConv2d can be described as:

[0050] y=(α w1 ⊙α f1 ⊙α c1 ⊙α s1 ⊙W1+…+α wn ⊙α fn ⊙α cn ⊙α sn ⊙W n )*x;

[0051] Among them, α wi Represents the convolution kernel w i The attention mark, α si ∈R k*k , They represent the attention along the spatial dimension, input channel dimension, and output channel dimension respectively.

[0052] Specifically, if Figure 5 (a) Position multiplication operation along the spatial dimension, for the convolution kernel w i , α si Assign different attention values ​​to the convolution parameters at the k*k spatial position; Figure 5 (b) Channel multiplication operation of input channel dimension, α ci Assign different attention values ​​to the convolution filters of different input channels; Figure 5 (c) Filter multiplication operation along the output channel dimension, α fi Assign different attention values ​​to the convolution filters of different output channels; Figure 5 (d) Kernel multiplication operation along the kernel dimension of the convolution kernel space, α wi Assign different values ​​to n overall convolution kernels.

[0053] These four types of attention are complementary and are implemented by progressively convolutional layers w i Multiplying different attention along the dimensions of position, channel, filter, and kernel will make the convolution operation have differences in various dimensions for the input, providing better performance to capture rich contextual information. Therefore, ODConv2d can greatly improve the feature extraction ability of convolution; more importantly, ODConv2d with fewer convolution kernels can achieve better performance.

[0054] The head part consists of four detection heads, each of which consists of a convolutional layer Conv and a convolutional layer Con2v. The upsampling of the SPPF module in the neck part is fused with the output of the shallow network in the backbone network and then convolved to output an additional shallow network detection head, adjusting the original three to four. Specifically, the detection head added to the shallow network is mainly used to detect small target objects, which improves the detection accuracy of small target objects and the detection accuracy of small drone targets with category label 0.

[0055] The present invention proposes to use SIOU as the loss function, and the SIOU formula is as follows:

[0056]

[0057] L SIoU is the SIOU loss value. On the basis of IoU loss, the following are added: shape loss Ω, which is the difference between the length and width ratio of the real detection box and the predicted detection box; Δ is the distance loss, which takes into account the distance between the real detection box and the predicted detection box, that is, the minimum circumscribed rectangular area of ​​the two; and then add IoU loss L IoU , 1-L IoU The non-overlapping parts of the predicted bounding boxes are emphasized. Finally, Focal Loss is performed on SIoU to prevent data imbalance and make the model focus on difficult-to-distinguish samples, thereby improving the training speed, reasoning accuracy, and model generalization.

[0058] In addition, based on SIOU, the present invention continues to improve the calculation of SIOU loss function, so that the model pays more attention to small target objects, so as to improve the detection accuracy of small UAV targets with category label 0.

[0059] Specifically, the loss value L calculated by the final SIOU loss function SIoU , adjust the adjustment coefficient L = γ·L SIoU, adjustment coefficient L = (Cw·h), where C is the maximum area of ​​the prediction detection box slightly larger than the set one, w is the width of the prediction detection box where the target object is located, h is the height of the prediction detection box where the target object is located, and w·h is the area of ​​the detection box where the target object is located. Therefore, the larger the prediction detection area where the target object is located, the smaller the loss, and the smaller the prediction detection box area where the target object is located, the greater the loss. The model will pay more attention to target objects with small areas, thereby improving the detection accuracy of small drone targets with category label 0.

[0060] In this embodiment of the present invention, the calculation formula of IoU is:

[0061] L IoU =(A∩B) / (A∪B);

[0062] That is, the result obtained by dividing the intersection of two areas by the union of the two areas. In practical applications, A is the predicted detection box (output by the improved YOLOv8 network model) and B is the real detection box (labeled detection box). In the training process of the target detection model, the IoU threshold is generally set to 0.5. When the ratio of the intersection area to the union area of ​​the two bounding boxes (the predicted box and the real box) is greater than 0.5, it is considered to be a true detection (true detection), otherwise it is a false detection (false detection). In this experiment (during model training), the IoU threshold was set to 0.8.

[0063] According to the training effect and performance of the network model, the hyperparameters are adjusted, and the training and parameter adjustment steps are repeated to obtain the best model. The established radar image dataset is input into the target detection model involved for training.

[0064] First, use the existing weights for a warm-up and training round, then verify with the validation data set, and output the evaluation results, including accuracy, recall, mAP70 value, F1 score, etc., and continuously adjust the hyperparameters, such as batch size, learning rate, etc., until the optimal hyperparameters and optimal model are obtained. Use the test set and various radar image data obtained in actual engineering applications, input them into the trained best model for prediction, and judge whether the output results meet the engineering requirements for the detection accuracy and efficiency of the target. In actual operation, both the model recognition accuracy and recognition speed are far superior to manual discrimination and traditional radar target detection algorithms, and can meet the needs of current engineering applications.

[0065] The present invention has been described exemplarily. Obviously, the specific implementation of the present invention is not limited to the above-mentioned method. As long as various non-substantial improvements are made using the method concept and technical solution of the present invention, or the concept and technical solution of the present invention are directly applied to other occasions without improvement, they are all within the protection scope of the present invention.

Claims

1. A radar image target detection method based on an improved YOLOv8 network model, characterized in that: The method comprises the following steps: S1, collect radar image samples, preprocess and enhance the collected radar image samples, and put them into a sample data set; S2, labeling the target objects in the radar image samples in the sample data set, and dividing the labeled sample data set into a training set and a test set; S3, training the improved YOLOv8 network model based on the training samples, and testing the trained improved YOLOv8 network model based on the test samples, until the recognition accuracy of the improved YOLOv8 network model reaches the required recognition accuracy; S4. Input the collected radar image into the trained improved YOLOv8 network model and output the target object in the radar image.

2. The radar image target detection method based on the improved YOLOv8 network model as claimed in claim 1, characterized in that: The radar image samples include restricted areas, medium-sized drones and small drones.

3. The radar image target detection method based on the improved YOLOv8 network model as claimed in claim 1, characterized in that: The improved YOLOv8 network model includes: It consists of backbone network, neck part and head part; Among them, the backbone network consists of a sequentially connected convolutional layer Conv, a dynamic convolution ODConv2d, a convolutional layer c2f, and an SPPF module.

4. The radar image target detection method based on the improved YOLOv8 network model as claimed in claim 3, characterized in that: The head part consists of four detection heads, each of which consists of a convolutional layer Conv and a convolutional layer Con2v.

5. The radar image target detection method based on the improved YOLOv8 network model as claimed in claim 1, characterized in that: The loss value of the improved YOLOv8 network model is L = γ·L SIoU , where L SIoU is the SIOU loss value, and γ is the adjustment coefficient.

6. The radar image target detection method based on the improved YOLOv8 network model as claimed in claim 5, characterized in that: Adjustment coefficient γ = (Cw·h), Among them, C is a constant slightly larger than the maximum area of ​​the set prediction detection box, w is the width of the prediction detection box where the target object is located, and h is the height of the prediction detection box where the target object is located.

7. The radar image target detection method based on the improved YOLOv8 network model as claimed in claim 5, characterized in that: SIOU loss value L SIoU The calculation formula is as follows: Among them, Ω is the shape loss, that is, the difference between the length and width ratios of the real detection box and the predicted detection box, Δ is the distance loss, that is, the difference between the minimum circumscribed rectangular area of ​​the real detection box and the predicted detection box; L IoU is the IoU loss.

8. The radar image target detection method based on the improved YOLOv8 network model as claimed in claim 1, characterized in that: The preprocessing process of radar image samples is as follows: The collected radar image samples are first converted into grayscale images using the OpenCV library, and then the radar image is subjected to local histogram equalization and then converted back to RGB images.

9. The radar image target detection method based on the improved YOLOv8 network model as claimed in claim 1, characterized in that: The radar image samples are enhanced by using at least one of the following methods: (a) Rotate the radar image based on the set rotation angle; (b) Flipping the radar image based on the set up, down, left, and right flip probabilities; (c) performing a radial transformation on the radar image based on a set affine transformation probability; (d) performing hybrid overlap on the radar images based on a set hybrid overlap probability; (e) Random mosaics of radar images.

Images