High-precision real-time target detection method

Abstract

The invention provides a high-precision real-time target detection method, and belongs to the technical field of target detection. The method comprises the following steps: taking a YOLOv8 framework as a model framework, and constructing a real-time target detection model in combination with a KAN network; carrying out pre-training on the real-time target detection model; and inputting an image to be detected into the pre-trained real-time target detection model, and outputting a target detection result. The KAN network has strong feature representation capability and can adapt to various types of target detection tasks. According to the method, the detection precision and the feature extraction capability are remarkably improved by introducing the KAN network. The KAN network uses a Kolmogorov-Arnold representation theory, so that the high-dimensional features of the image can be efficiently extracted. In combination with the real-time performance of the YOLOv8, the whole system can still keep higher detection speed while keeping high precision by optimizing the network design.

Description

Technical Field

[0001] The present invention relates to the technical field of target detection, and in particular to a high-precision real-time target detection method. Background Art

[0002] Object detection is one of the core tasks in the field of computer vision, aiming to identify and locate multiple object categories and their spatial locations in images or videos. Its key technologies can be divided into region-based detection methods (such as the R-CNN series) and regression-based detection methods (such as the YOLO series). Region-based methods first generate candidate regions (Region Proposal), and then classify and regress the candidate regions through a convolutional neural network (CNN); while regression-based methods directly predict the category and location of the target through an end-to-end network, significantly improving detection speed. These technologies are widely used in scenarios such as autonomous driving, video surveillance, and robot navigation, and are particularly crucial in the field of autonomous driving, which requires extremely high real-time performance and accuracy.

[0003] Among the current mainstream detection methods, region-based detection frameworks (such as Faster R-CNN and Mask R-CNN) generate candidate boxes through region proposal networks and combine them with a two-stage classification and regression strategy to achieve high-precision detection. However, their high computational complexity makes them difficult to meet real-time requirements. Single-stage regression-based methods (such as YOLO and SSD) have real-time advantages, but because they directly regress target attributes, they are prone to missed detections or positioning errors in dense and small target detection, and the feature extraction process may lose spatial details. Furthermore, both methods rely on the feature representation capabilities of traditional CNNs, which still have limitations in their robustness to complex scenes and their ability to integrate multi-scale features.

[0004] Existing methods face the following key issues: (1) Region-based methods have high computational complexity and slow detection speed due to candidate region generation and multiple feature calculations; (2) Regression-based methods are prone to losing spatial information during feature extraction due to dimensionality reduction and pooling operations, resulting in insufficient detection accuracy for small and dense targets; (3) Traditional CNNs have limited feature representation capabilities and are unable to fully model the diversity of targets and the interference of complex backgrounds; (4) Existing frameworks are not robust enough to dynamic scenes such as lighting changes and occlusions, which affects the stability of practical applications. How to balance speed and accuracy and improve feature expression capabilities and robustness has become a core challenge in the field of multi-target detection.

[0005] Therefore, a high-precision real-time target detection method is needed. Summary of the Invention

[0006] In view of this, the present invention provides a high-precision real-time target detection method. By combining the efficient feature representation capability of the KAN network (Kolmogorov-Arnold Networks) and the real-time detection capability of the YOLO series, a multi-target detection method with low computational complexity, high detection accuracy, strong feature representation capability and good robustness is provided.

[0007] To this end, the present invention provides the following technical solutions:

[0008] A high-precision real-time target detection method, comprising:

[0009] Using the YOLOv8 framework as the model framework and combining it with the KAN network to build a real-time target detection model;

[0010] Pre-training the real-time object detection model;

[0011] The image to be tested is input into the pre-trained real-time target detection model and the target detection result is output.

[0012] Furthermore, the real-time target detection model includes:

[0013] The target backbone network optimized by the KAN network, the target neck network optimized by the KAN network, and the target head network optimized by the KAN network;

[0014] The target backbone network extracts multi-scale features of the image;

[0015] The target neck network fuses features at different scales;

[0016] The target head network predicts object category and location.

[0017] Furthermore, the target backbone network includes:

[0018] The target conv module optimized by the KAN network, the KAN network module, the target c2f module optimized by the KAN network, and the target SPPF module optimized by the KAN network.

[0019] Furthermore, the target neck network includes:

[0020] Wavelet filtering module and target c2f module.

[0021] Furthermore, the data processing process of the target conv module includes:

[0022] Target conv module output data

[0023] =SiLU(BN(KANConv2DLayer(target conv module input data)))

[0024] Among them, SiLU represents the SiLU activation function; BN represents normalization processing; KANConv2DLayer represents the KAN two-dimensional convolutional layer.

[0025] Furthermore, the data processing process of the target head network includes:

[0026] Head_kan=KAN(Conv2d(ConvK(ConvK(target head network input data))));

[0027] Among them, ConvK represents the target conv module calculation process, Conv2d represents 2D convolution; Head_kan

[0028] Indicates the target detection result.

[0029] Furthermore, the data processing process of the target c2f module includes:

[0030] Target c2f module output data =

[0031] Conv(Concat(Bottleneck_kan(Bottleneck_kan(...Conv(target c2f module input data))),..;

[0032] Bottleneck k an=

[0033] ADD(ConvK(ConvK(target c2f module input data)),target c2f module input data);

[0034] Among them, Conv means convolution; Concat means splicing; ADD means addition.

[0035] Furthermore, the real-time target detection model pre-training includes:

[0036] Collect image datasets;

[0037] Convert image data into YOLO format and perform preprocessing to obtain preprocessed data;

[0038] Feed the preprocessed data into the real-time object detection model for learning.

[0039] Furthermore, the preprocessing includes:

[0040] Resize the image to a preset size;

[0041] Data augmentation through Albumentations.

[0042] Advantages and positive effects of the present invention:

[0043] The present invention significantly improves detection accuracy and feature extraction capabilities by introducing the KAN network. The KAN network uses the Kolmogorov-Arnold representation theory to efficiently extract high-dimensional features of images.

[0044] The network uses a learnable activation function at the edge of the node, replacing the fixed activation function design in MLP, which can better fit the data. Considering the problem that convolutional networks sometimes have insufficient feature extraction, KAN

[0045] The network is used for feature extraction, which can effectively improve the model's feature extraction capabilities.

[0046] By combining the real-time performance of YOLOv8, this paper optimizes the network design, enabling the entire system to maintain high accuracy while maintaining a high detection speed. The KAN network design makes the feature extraction process more efficient and reduces the consumption of computing resources. Combined with the lightweight design of YOLOv8, the entire system can still operate efficiently even with limited hardware resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0048] Figure 1 This is a flow chart of the high-precision real-time target detection method in Example 1 of the present invention;

[0049] Figure 2 This is a structural diagram of a real-time target detection model in Example 1 of the present invention;

[0050] Figure 3 This is a structural diagram of the ConvK module in Example 1 of the present invention;

[0051] Figure 4 This is a structural diagram of the C2f_kan module in Example 1 of the present invention;

[0052] Figure 5 This is a structural diagram of the Bottleneck-kan module in Example 1 of the present invention;

[0053] Figure 6 This is a structural diagram of the SPPF_kan module in Example 1 of the present invention;

[0054] Figure 7This is a structural diagram of the Head_kan module in Example 1 of the present invention;

[0055] Figure 8 This is a flow chart of target detection model training in Example 2 of the present invention. DETAILED DESCRIPTION

[0056] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0057] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0058] The present invention provides a high-precision real-time target detection method, combining Figure 1 Flowchart, the steps of this method include: using the YOLOv8 framework as the model framework and combining it with the KAN network to build a real-time target detection model; inputting the data to be detected into the real-time target detection model and outputting the target detection results.

[0059] Example 1

[0060] Combine Figure 2 As shown, in this embodiment, the YOLOv8 model framework is combined with the KAN network to build a high-precision real-time target detection model; the real-time target detection model includes:

[0061] The target backbone network optimized by the KAN network, the target neck network optimized by the KAN network, and the target head network optimized by the KAN network; among them, the target backbone network extracts multi-scale features of the image; the target neck network fuses features of different scales; the target head network predicts object category and position.

[0062] Target backbone network, including:

[0063] The target conv module optimized by the KAN network is denoted as the ConvK module, and its structure is as follows Figure 3 As shown, a KAN network layer is added before the convolutional layer;

[0064] KAN network module;

[0065] The target c2f module optimized by the KAN network is denoted as C2f_kan module, and its structure is as follows Figure 4 As shown in Figure 1, a KAN network layer is added after the Bottleneck layer to construct the target Bottleneck layer optimized by the KAN network, which is denoted as Bottleneck-kan.

[0066] Among them, the target Bottleneck layer (Bottleneck-kan) structure is as follows Figure 5 As shown;

[0067] The target SPPF module optimized by the KAN network is recorded as the SPPF-kan module, and its structure is as follows: Figure 6 As shown in Figure 2, the convolution operation is implemented through the target conv module.

[0068] Targeted neck networks, including:

[0069] Through the wavelet filtering module and the target c2f module optimized by the KAN network, the structure is as follows Figure 4 shown.

[0070] Target head network, including:

[0071] The target Head network optimized by the KAN network is denoted as the Head_kan module, and its structure is as follows: Figure 7 shown.

[0072] Example 2

[0073] Data processing of real-time object detection model:

[0074] Step 1: The target backbone network extracts multi-scale features of the image:

[0075] S1, preprocess the data X orig Send it to the ConvK module to extract features and output the first feature data;

[0076] The ConvK module is the convolution implementation of the KAN network. The convolution kernel of the convolution process is replaced with a learnable activation function: first, it passes through the KAN convolution module, then passes through the normalized BN module, and finally passes through a SiLU activation function to obtain the first feature data X1; the structure is as follows Figure 3 As shown, the formula is expressed as:

[0077] X1=SiLU(BN(KANConv2DLayer(X orig )))

[0078] Among them, X1 represents the output feature data of the ConvK module, SiLU represents the SiLU activation function, BN represents the normalization processing, and KANConv2DLayer represents the KAN two-dimensional convolution layer.

[0079] S2. Send the first feature data X1 to the KAN module. The KAN module is a KAN network. The KAN network uses the B-spline function to fit the characteristics of the data. At the same time, it can also change the dimension and size of the data to better learn the features. After passing through this module, the second feature data X2 is obtained. The formula is expressed as:

[0080] X2=KAN(X1)

[0081] Wherein, X2 represents the second feature data output by the KAN module.

[0082] S3. Input the second feature data into the C2f_kan module:

[0083] The C2f_kan module is a combination of the C2f block and the KAN convolution. Adding the KAN convolution to the C2f module can improve the feature extraction capability of the C2f_kan module while reducing the number of parameters. The structure is as follows Figure 4 As shown, the formula is expressed as:

[0084] X3=Conv(Concat(Bottleneck_kan(Bottleneck_kan(...Conv(X2)))),

[0085] ...,Bottleneck_kan(Conv(X2))))

[0086] Among them, X3 represents the third feature data output by the C2f_kan module; Bottleneck_kan=ADD(ConvK(ConvK(X2)),X2); The structure diagram of Bottleneck_kan is as follows Figure 5 shown.

[0087] The KAN network is added to the Bottleneck block, which improves the feature extraction capability of the original Bottleneck block and reduces the number of parameters of the Bottleneck block.

[0088] S4, the third feature data X3 is input into the KAN module, and the fourth feature data X4 is output;

[0089] The fourth feature data X4 is input into the C2f_kan module, which outputs the fifth feature data X5;

[0090] The fifth feature data X5 is input into the KAN module, and the sixth feature data X6 is output.

[0091] S5, the sixth feature data X6 is input into the SPPF_kan module and the seventh feature data X7 is output. The structure of the SPPF_kan module is as follows: Figure 6 As shown, the formula is expressed as:

[0092] X7=ConvK(Concat(MaxP2d(MaxP2d(MaxP2d(ConvK(X b )))),

[0093] MaxP2d(MaxP2d(ConvK(X6))),MaxP2d(ConvK(X6))))

[0094] Among them, X7 represents the output of the seventh feature data, ConvK represents the target conv module calculation process, MaxP2d represents the two-dimensional maximum pooling, and X6 represents the output of the sixth feature data.

[0095] Step 2: The target neck network fuses features of different scales:

[0096] S6, the seventh feature data X7 is input into the WaveUpPool module, and the eighth feature data X8 is output. The formula is expressed as:

[0097] X8=WaveUpPool(X7)

[0098] Among them, the WaveUpPool module is a wavelet filter.

[0099] S7, input the eighth feature data X8 and the fifth feature data X5 into the connection module for splicing to obtain the ninth feature data X9, which is expressed as follows:

[0100] X9=Concat(X8,X5)

[0101] S8, input the ninth feature data X9 into the C2f_kan module to obtain the tenth feature data X 10 ; Tenth feature data X 10 After the WaveUpPool module, the eleventh characteristic data X is obtained 11 ; Among them, the structure and function of the C2f_kan module are the same as those in step S3, and the structure and function of the WaveUpPool module are the same as those in S7;

[0102] The eleventh feature data X 11The third characteristic data X3 is input into the connection module for splicing to obtain the twelfth characteristic data X 12 , the formula is:

[0103] X 12 =Concat(X 11 ,X3)

[0104] S9, twelfth characteristic data X 12 After the C2f_kan module, the thirteenth feature data X is obtained 13 ;

[0105] The thirteenth feature data X 13 After the KAN module, the fourteenth feature data X is obtained 14 .

[0106] S10, the fourteenth data X 14 and the tenth feature data X 10 Input the connection module and perform splicing to obtain the fifteenth feature data X 15 , the formula is:

[0107] X 15 =Concat(X 14 ,X 10 )

[0108] Fifteenth characteristic data X 15 After the C2f_kan module, the sixteenth feature data X is obtained 16 , and then pass through the KAN module to obtain the seventeenth feature data X 17 .

[0109] S11, the seventeenth feature data X 17 Combined with the seventh feature data X7, the eighteenth feature data X 18 , the formula is:

[0110] X 18 =Concat(X 17 ,X7)

[0111] Eighteenth characteristic data X 18 After the C2f_kan module, the nineteenth feature data X is obtained 19 , and then pass through the KAN module to obtain the twentieth feature data X 20 .

[0112] Step 3: The target head network predicts the object category and location:

[0113] S12. Input the fourteenth feature into the Head_kan module to obtain the first output Y1, which is expressed as:

[0114] Y1=Head_kan(X 14 )

[0115] Among them, the Head_kan module is the regression loss, and its structure is as follows Figure 7 As shown, the formula is expressed as:

[0116] Head_kan=KAN(Conv2d(ConvK(ConvK(X 14 )))).

[0117] S13. Input the seventeenth feature data into the Head_kan module to obtain the second output Y2.

[0118] Y2=Head_kan(X 17 )

[0119] S14, the twentieth feature data X 20 Input the Head_kan module to get the third output Y3.

[0120] Y3=Head_kan(X 20 )

[0121] S15. Finally, the best prediction result is obtained through non-maximum suppression.

[0122] This method converts data into the YOLO format and then preprocesses it, including resizing and augmenting it with Albumentations. This preprocessed data is then fed into the model for feature learning. After training, the image to be tested is fed into the trained model to generate test results, effectively extracting high-dimensional features from the image.

[0123] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A high-precision real-time target detection method, characterized in that: include: Using the YOLOv8 framework as the model framework and combining it with the KAN network to build a real-time target detection model; Pre-training the real-time object detection model; The image to be tested is input into the pre-trained real-time target detection model and the target detection result is output.

2. A high-precision real-time target detection method according to claim 1, characterized in that: The real-time target detection model includes: The target backbone network optimized by the KAN network, the target neck network optimized by the KAN network, and the target head network optimized by the KAN network; The target backbone network extracts multi-scale features of the image; The target neck network fuses features at different scales; The target head network predicts object category and location.

3. A high-precision real-time target detection method according to claim 2, characterized in that: The target backbone network includes: The target conv module optimized by the KAN network, the KAN network module, the target c2f module optimized by the KAN network, and the target SPPF module optimized by the KAN network.

4. A high-precision real-time target detection method according to claim 2, characterized in that: The target neck network includes: Wavelet filtering module and target c2f module.

5. A high-precision real-time target detection method according to claim 3, characterized in that: The data processing process of the target conv module includes: Target conv module output data =SiLU(BN(KANConv2DLayer(target conv module input data))) Among them, SiLU represents the SiLU activation function; BN represents normalization processing; KANConv2DLayer represents the KAN two-dimensional convolutional layer.

6. A high-precision real-time target detection method according to claim 2, characterized in that: The data processing process of the target head network includes: Head_kan=KAN(Conv2d(ConvK(ConvK(target head network input data)))); Among them, ConvK represents the target conv module calculation process, Conv2d represents 2D convolution; Head_kan represents the target detection result.

7. A high-precision real-time target detection method according to claim 3, characterized in that: The data processing process of the target c2f module includes: Target c2f module output data = Conv(Concat(Bottleneck_kan(Bottleneck_kan(...Conv(target c2f module input data))),..; Bottleneck k an= ADD(ConvK(ConvK(target c2f module input data)),target c2f module input data); Among them, Conv means convolution; Concat means splicing; ADD means addition.

8. A high-precision real-time target detection method according to claim 1, characterized in that: The real-time target detection model pre-training includes: Collect image datasets; Convert image data into YOLO format and perform preprocessing to obtain preprocessed data; Feed the preprocessed data into the real-time object detection model for learning.

9. A high-precision real-time target detection method according to claim 8, characterized in that: The pretreatment includes: Resize the image to a preset size; Data augmentation through Albumentations.

Images