Road self-patrol evaluation system based on YOLOv9

Abstract

The invention discloses a road self-patrol evaluation system based on YOLOv9, and relates to the technical field of road detection, the road self-patrol evaluation system comprises a camera, a positioning module, a radar, a main control unit and a target algorithm detection module, the camera, the positioning module, the radar and the main control unit are carried on a road cruise vehicle, and the target algorithm detection module is arranged on a user operation PC; the camera is used for automatically shooting road images and transmitting the road images to the user operation PC through the local area network for algorithm analysis to judge whether cracks and potholes exist or not. By fusing the Internet of Things and visual intelligent detection, intelligent inspection and evaluation of road conditions are realized, the efficiency and the accuracy are improved, meanwhile, personnel can be prevented from entering a dangerous area by utilizing machine automatic inspection, and the safety risk is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of road detection, and in particular to a road self-inspection system based on YOLOv9. Background Art

[0002] In modern society, roads, as an important part of urban infrastructure, play a vital role in economic development and people's lives. Good road conditions are the premise for ensuring traffic safety and improving traffic efficiency. However, due to environmental factors, vehicle load and material aging, cracks, potholes and other damages will appear on the road. Regular inspection and maintenance are essential to maintain road performance and extend its service life. Traditional road inspection relies on manual visual inspection or simple instrument measurement. This method is labor-intensive, costly, and has a long cycle. It is also easily affected by subjective factors, resulting in detection efficiency and accuracy that are difficult to meet the needs of modern urban development. However, with the development of science and technology and the advent of the computer age, the pixel of camera equipment and computer computing power have been greatly improved, which helps to record images of the road surface and further use computer equipment to complete the automatic identification of road diseases. This operation mode has many advantages, such as no need for on-site operations, no need to close traffic, high inspection efficiency, high accuracy and can greatly reduce the total cost of road disease inspection.

[0003] While hardware technologies such as camera equipment and computers are developing, relevant software algorithms have also made great progress. As early as the 1980s, scholars began to apply digital image processing technology to the intelligent extraction of road surface diseases and achieved good results. With the development of computer automatic control, the road crack inspection system based on computer automatic control has also developed rapidly, and scientific research institutions in many countries have developed intelligent road inspection vehicles. Although domestic and foreign scholars have achieved many results in this field, the automatic detection system of crack diseases based on digital images has yet to be developed and improved, and some key issues have not been completely resolved, such as data acquisition and transmission, crack identification and positioning, and display of data processing results. These key issues determine the safety and reliability of road construction and play a vital role in extending the service life of roads.

[0004] In addition to the problems in road surface disease inspection during the operation and maintenance phase, there are also measures to be improved in other stages of the road engineering quality management system. Therefore, it is particularly urgent to develop an efficient, automated, and intelligent road self-inspection system. Summary of the invention

[0005] The purpose of the present invention is to provide a road self-inspection system based on YOLOv9 to solve the problems raised in the above background technology.

[0006] To solve the above problems, the technical solutions adopted by the present invention are as follows:

[0007] The road self-inspection system based on YOLOv9 includes

[0008] A camera, a positioning module, a radar, a main control unit and a target algorithm detection module, wherein the camera, the positioning module, the radar and the main control unit are mounted on the road cruise vehicle, and the target algorithm detection module is set on a user-operated PC;

[0009] The camera is used to automatically take pictures of the road and transmit them to the user's PC through the local area network for algorithm analysis to determine whether there are cracks or potholes;

[0010] When the target algorithm detection module detects road damage in the image taken by the camera, it reads the positioning module data and sends the information to the EMQX server built on the Tencent Cloud lightweight application server through the ESP chip and further uploads it to the user interface, while displaying the image and detection results to the user.

[0011] Preferably, the main control unit is a Luban Cat microcomputer equipped with a Linux system and capable of visualizing the state of the cruise vehicle and remotely controlling the cruise vehicle through a local area network and a virtual machine.

[0012] Preferably, the sensors used in the positioning module are odom, GPS and IMU.

[0013] Preferably, the ESP chip packages the collected data and sends it to the cloud EMQX server, performs data analysis in the cloud, and displays real-time working information on the PC.

[0014] Preferably, the target algorithm detection module adopts the YOLOv9 target detection algorithm, integrates the programmable gradient information PGI and the generative potential embedding GLEAN architecture for object detection into YOLOv9, further enhances its performance in the object detection task, and forms the programmable gradient information PGI and its network architecture method;

[0015] The programmable gradient information PGI and its network architecture method include the following four different network designs:

[0016] (1) PAN: used to improve feature fusion to improve the performance of object detection;

[0017] (2) RevCo l: A network design used to reduce information loss by maintaining information flow without loss through a reversible column structure;

[0018] (3) Deep supervision: improving learning efficiency and final model performance by inserting additional supervision signals at multiple layers of the network;

[0019] (4) Programmable Gradient Information (PGI): used for forward propagation without skip links.

[0020] Preferably, the programmable gradient information PGI includes a main branch, an auxiliary reversible branch and multi-level auxiliary information;

[0021] Wherein, the main branch is used for the architecture of reasoning;

[0022] The auxiliary reversible branch is used to generate reliable gradients to supply the main branch for back propagation;

[0023] The multi-level auxiliary information is used to control the main branch to learn the plannable multi-level semantic information.

[0024] Preferably, the YOLOv9 target detection algorithm includes a detection model, a training model, a model evaluation, a model lightweight, a model deployment and a detection effect;

[0025] The detection models include three models: YOLOv9s, YOLOv9m, and YOLOv9 l. The difference between them is the different parameter amounts. YOLOv9s is the model with the smallest parameter amount. The depth and width of the three models are successively deepened, and the accuracy is also successively improved;

[0026] The model lightweighting is achieved by quantizing the model and compressing the model;

[0027] The model deployment is used to apply the trained model to practical applications, including task encapsulation and integration, platform adaptation and optimization, to provide users with efficient and accurate image recognition and detection services.

[0028] Preferably, the YOLOv9s network structure consists of four parts, including an input end, a backbone feature extraction network Backbone, a bottleneck layer Neck, and a detection module output layer Head.

[0029] Preferably, the backbone feature extraction network Backbone includes a Focus module, a convolution module Conv, a C3 module and an SPP module;

[0030] The Focus module is used to slice the image;

[0031] The convolution module Conv module integrates three operations: convolution, normalization and activation function, and the activation function is the ReLU function;

[0032] The C3 module includes three convolutions and several bottleneck modules. The bottleneck modules use 1×1 and 3×3 convolutions to extract features and use residual connections to enhance the network feature extraction capability.

[0033] The SPP module is also called spatial pyramid pooling, which improves the perception field of the feature map by performing pooling operations on the feature map through pooling kernels of different sizes and then concat.

[0034] Preferably, the bottleneck layer Neck adopts a bottom-up connection structure of a feature pyramid, and the detection module output layer Head includes several detection heads, which respectively predict feature maps of different sizes, corresponding to large targets, medium targets and small targets. The large-size feature maps are used to predict small targets, and the small-size feature maps are used to predict large targets. Each grid of the feature map corresponds to 3 anchor sizes with different width-to-height ratios. At the same time, prediction and regression are performed in the channel dimension of the feature map through the target position and classification information of the anchor prior box.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] The present invention integrates the Internet of Things and visual intelligent detection to achieve intelligent inspection and evaluation of road conditions, improve efficiency and accuracy, and use automatic machine inspections to prevent people from entering dangerous areas and reduce safety risks.

[0037] The present invention utilizes multi-dimensional data collection and analysis and multi-sensor collaboration to provide decision support through data processing and analysis, monitor road conditions around the clock and in all directions, and quickly provide comprehensive evaluation results to improve the service efficiency and profits of road administration departments and engineering companies.

[0038] The PGI designed by the present invention solves the problem that deep monitoring can only be used for extremely deep neural network architectures, so that the new lightweight architecture can be truly applied to daily life. It has strong flexibility and scalability. In addition to the inspection vehicle test, the project also completed the drone test, modified the drone, and integrated other sensors such as cameras and GPS into the system. In the future, it can be expanded to other application fields that require drone testing.

[0039] This paper theoretically analyzes the existing deep neural network architecture from the perspective of reversible functions, and successfully explains many phenomena that were difficult to explain in the past through this process. On this basis, we also designed PGI and auxiliary reversible branches, and achieved good results.

[0040] The present invention uses intelligent Internet of Things technology to transmit image data to cloud computers in real time for processing and evaluation, and quickly generates evaluation reports and suggestions. It provides a simple graphical operation interface to facilitate users to configure inspection tasks and adjust image processing algorithm parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1This is a schematic diagram of the execution of the road self-inspection system based on YOLOv9;

[0042] Figure 2 This is the YOLOv9 architecture diagram of the road self-inspection system based on YOLOv9;

[0043] Figure 3 A schematic diagram of the programmable gradient information PGI and its network architecture method of the present invention;

[0044] Figure 4 It is a feature extraction network flow chart of the present invention;

[0045] Figure 5 Schematic diagram of three training files of YOLOv9 of the present invention;

[0046] Figure 6 It is a schematic diagram of the data set format of the present invention;

[0047] Figure 7 This is a schematic diagram of the fall.yaml file content of the present invention;

[0048] Figure 8 Schematic diagram of YOLOv9-C and YOLOv9-E of the present invention;

[0049] Fig. 9 It is a schematic diagram of the pre-training weights of YOLOv9-C of the present invention;

[0050] Fig.10 It is a schematic diagram of specific parameters of the present invention;

[0051] Fig.11 It is a schematic diagram of the operation result of the present invention;

[0052] Fig.12 It is a schematic diagram of a specific storage path of the present invention;

[0053] Fig.13 Schematic diagram of comparative training parameters and indicators of the most advanced real-time target detector of the present invention. DETAILED DESCRIPTION

[0054] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments 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 creative work are within the scope of protection of the present invention.

[0055] Example:

[0056] See also Figure 1-Figure 13As shown, the present invention is a road self-inspection system based on YOLOv9, including

[0057] The camera, positioning module, radar, main control unit and target algorithm detection module are mounted on the road cruiser, and the target algorithm detection module is set on the user-operated PC;

[0058] The camera is used to automatically take pictures of the road and transmit them to the user's PC through the local area network for algorithm analysis to determine whether there are cracks or potholes.

[0059] When the target algorithm detection module detects road damage in the image taken by the camera, it reads the positioning module data and sends the information to the EMQX server built on the Tencent Cloud lightweight application server through the ESP chip and further uploads it to the user interface. At the same time, the image and detection results are displayed to the user.

[0060] The main control unit is a Luban Cat microcomputer, which is equipped with a Linux system and can visualize the cruise vehicle status and remotely control the cruise vehicle through a local area network and a virtual machine.

[0061] The sensors used in the positioning module are odom, GPS and IMU.

[0062] The ESP chip packages the collected data and sends it to the EMQX server in the cloud, where it performs data analysis and displays real-time working information on the PC;

[0063] The ESP series chips can not only efficiently collect and transmit data from various systems, but also have the characteristics of being easy to use, quick to get started, and affordable;

[0064] The EQMX message server has the ability to stably carry large-scale MQTT client connections, with a single server node supporting 2 million connections; distributed node clusters, fast and low-latency message routing; expansion within the message server, support for customization of multiple authentication methods, efficient storage of messages to the backend database and complete IoT protocol support.

[0065] The target algorithm detection module adopts the YOLOv9 target detection algorithm, integrates the programmable gradient information PGI and the generative latent embedding GLEAN architecture for object detection into YOLOv9, further enhances its performance in object detection tasks, and forms the programmable gradient information PGI and its network architecture method;

[0066] Programmable Gradient Information (PGI) and its network architecture approach include the following four different network designs:

[0067] (1) PAN: used to improve feature fusion to improve the performance of object detection;

[0068] (2) RevCol: A network design used to reduce information loss by maintaining information flow without loss through a reversible column structure;

[0069] (3) Deep supervision: improving learning efficiency and final model performance by inserting additional supervision signals at multiple layers of the network;

[0070] (4) Programmable Gradient Information (PGI): used for forward propagation without skip links.

[0071] As can be seen above, the YOLO series has long revolutionized the field of object detection by introducing breakthrough concepts in computer vision, such as processing the entire image at once through convolutional neural networks (CNNs). From YOLOv1 to the latest YOLOv9, each of its iterations continues to improve and integrate advanced technologies to improve accuracy, speed, and efficiency, making it the preferred solution for real-time target detection across fields and scenarios;

[0072] Specifically, YOLOv9 is the latest version of the YOLO (You Only Look Once) series of real-time target detection systems. It builds on previous versions and incorporates advances in deep learning technology and architectural design to achieve superior performance in object detection tasks. YOLOv9 combines the concept of Programmable Gradient Information (PGI) with the General ELAN (GELAN) architecture, representing a major leap in accuracy, speed, and efficiency.

[0073] The purpose of PGI is to assist reversible branching (such as Figure 3 The information bottleneck problem is solved by using the dotted box in the figure to provide more reliable gradients for deep networks without increasing the inference cost. With this design, even lightweight and shallow neural networks can achieve effective information retention and accurate gradient updates. Figure 3 The main branch shown in , through the reliable gradient information provided by the auxiliary reversible branch, can obtain more effective target task features without losing important information due to the information bottleneck.

[0074] The programmable gradient information PGI includes a main branch, an auxiliary reversible branch, and multi-level auxiliary information;

[0075] Among them, the main branch is used for the reasoning architecture;

[0076] The auxiliary reversible branch is used to generate reliable gradients to supply the main branch for back propagation;

[0077] Multi-level auxiliary information is used to control the main branch to learn plannable multi-level semantic information.

[0078] YOLOv9 target detection algorithm includes detection model, training model, model evaluation, model lightweighting, model deployment and detection effect;

[0079] The detection models include YOLOv9s, YOLOv9m, and YOLOv9l. The difference between them is the number of parameters. YOLOv9s is the model with the smallest number of parameters. The depth and width of the three models increase successively, and the accuracy also increases successively.

[0080] Model lightweighting is achieved by quantizing and compressing the model;

[0081] Among them, the model training of YOLOv9 in this project is quite special. Its detection head is divided into several different training methods. At the same time, because it has just been released, Figure 5 You can see that YOLOv9 gives three training files corresponding to:

[0082] train: main branch;

[0083] train_dual: one auxiliary branch + one main branch;

[0084] triple_branch: 2 auxiliary branches + 1 main branch.

[0085] Training the model involves the following steps:

[0086] 1. Dataset preparation, see Figure 6 As shown;

[0087] The fall.yaml file content is as follows Figure 7 As shown, the path, category and number of categories need to be modified according to your own needs:

[0088] 2. Download pre-trained weights, such as Figure 8 As shown;

[0089] 3. Training, if you load the pre-trained weights of YOLOv9-C, you need to modify yolov9-c.yaml, where the pre-trained weights of YOLOv9-C are as follows Fig. 9 As shown;

[0090] 4. Modify the parameters in train_dual.py. For specific parameters, see Fig.10 As shown;

[0091] 5. Run and get the running results. The running results are shown in Fig.11 As shown;

[0092] 6. Training is finished and saved in Fig.12 in the path.

[0093] Model evaluation, testing and evaluating the trained model through the test data set, Fig.13 The comparison of our proposed YOLOv9 with other real-time object detectors trained from scratch is listed. Overall, the best performing methods among the existing methods are the lightweight model YOLOMS-S[7], the medium model YOLOMS[7], the general model YOLOv7AF

[63] , and the large model YOLOv8-X

[15] . Compared with the lightweight and medium models YOLOMS[7], YOLOv9 has about 10% fewer parameters and 5-15% less computation, but still has an improvement of 0.4%-0.6% in AP. Compared with YOLOv7AF, YOLOv9-C has 42% fewer parameters and 21% less computation, but achieves the same AP (53%). Compared with YOLOv8-X, YOLOv9-X has 15% fewer parameters and 25% less computation, and has a significant 1.7% improvement in AP. The above comparison results show that compared with the existing methods, our proposed YOLOv9 has obvious improvements in all aspects;

[0094] In addition, we also include the ImageNet pre-trained model in the comparison and the results are shown in Fig.13 We compare based on parameters and computational effort. In terms of the number of parameters, the best performing large model is RTDETR

[43] . Figure 5 As can be seen in Figure 1, YOLOv9 using traditional convolutions performs even better in parameter utilization than YOLOMS using deep convolutions. For parameter utilization of large models, it also far exceeds RTDETR using the ImageNet pre-trained model. Even better, in deep models, YOLOv9 shows a huge advantage in using PGI. By accurately retaining and extracting the information needed to map data to the target, our method only requires 64% of the parameters to maintain the same accuracy as RTDETR-X.

[0095] In terms of computational complexity, the best existing models, from smallest to largest, are YOLOMS[7], PPYOLOE

[74] , and RTDETR

[43] . Figure 5 As can be seen from the figure, YOLOv9 is far superior to methods trained from scratch in terms of computational complexity. In addition, if compared with methods based on deep convolution and ImageNet pre-trained models, YOLOv9 is also very competitive.

[0096] Depend on Fig.13 It can be seen that compared with the lightweight and medium-sized models YOLOMS: the parameters are reduced by about 10%, the calculation amount is reduced by 5% to 15%, but the AP is still improved by 0.4% to 0.6%;

[0097] Compared with YOLOv7AF: YOLOv9-c has 42% fewer parameters and 21% less computation, but achieves the same AP (53%);

[0098] Compared with YOLOv8-x: YOLOv9-x has 15% fewer parameters, 25% less computation, and a significant 1.7% improvement in AP.

[0099] Model deployment is used to apply trained models to practical applications, including task encapsulation and integration, platform adaptation and optimization, to provide users with efficient and accurate image recognition and detection services.

[0100] The YOLOv9s network structure consists of four parts, including the input end, the backbone feature extraction network Backbone, the bottleneck layer Neck, and the detection module output layer Head.

[0101] The backbone feature extraction network Backbone includes Focus module, Conv module, C3 module and SPP module;

[0102] The Focus module is used to slice images;

[0103] The convolution module Conv module integrates three operations: convolution, normalization and activation function. The activation function is the ReLU function.

[0104] The C3 module contains three convolutions and several bottleneck modules. The bottleneck modules use 1×1 and 3×3 convolutions for feature extraction and residual connections to enhance the network’s feature extraction capabilities.

[0105] The SPP module is also called spatial pyramid pooling. It improves the perception field of the feature map by pooling the feature map with pooling kernels of different sizes and then concat.

[0106] The bottleneck layer Neck adopts a bottom-up connection structure of the feature pyramid. The output layer Head of the detection module includes several detection heads, which predict feature maps of different sizes, corresponding to large targets, medium targets and small targets. The large-size feature maps are used to predict small targets, and the small-size feature maps are used to predict large targets. Each grid of the feature map corresponds to three anchor sizes with different width-to-height ratios. At the same time, prediction and regression are performed in the channel dimension of the feature map through the target position and classification information of the anchor prior box.

[0107] The camera uses a binocular camera;

[0108] As can be seen from the above, the cruise car uses the microcomputer on the car as the main control unit, runs the Linux-based ROSKineticKame robot operating system, integrates radar, binocular camera, GPS and IMU sensors, and realizes the collaborative work of multiple sensors through the software architecture built by ROS. The system uses laser radar to build maps, and after the map is built, it uses algorithms to plan global path cruise, and uses binocular cameras to detect small targets in a dynamic shooting environment;

[0109] When cruising on the road, ESP collects data from modules such as the odometer, GPS, and IMU and uploads the data to the local area network. The images captured by the IP camera are also displayed on the user interface via the local area network and identified and judged. Users can view the location information, images, and detection results of road hazards on the interface, and can monitor and remotely control the cruise vehicle in real time through ROSRviz.

[0110] The standard parts used in the present invention can all be purchased from the market, and the special-shaped parts can be customized according to the description and the drawings. The specific connection methods of each part adopt conventional means such as mature bolts, rivets, welding, etc. in the prior art. The machinery, parts and equipment all adopt conventional models in the prior art, and the circuit connection adopts the conventional connection method in the prior art, which will not be described in detail here. The contents not described in detail in this specification belong to the prior art known to professional and technical personnel in this field.

[0111] In the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0112] In the present invention, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may mean that the first and second features are in direct contact, or the first and second features are in indirect contact through an intermediate medium. Moreover, a first feature being "above", "above" or "above" a second feature may mean that the first feature is directly above or obliquely above the second feature, or simply means that the first feature is higher in level than the second feature. A first feature being "below", "below" or "below" a second feature may mean that the first feature is directly below or obliquely below the second feature, or simply means that the first feature is lower in level than the second feature.

[0113] In the drawings of the embodiments disclosed in the present invention, only the structures related to the embodiments disclosed in the present invention are involved, and other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of the present invention can be combined with each other.

[0114] Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. The road self-inspection system based on YOLOv9 is characterized by: It includes a camera, a positioning module, a radar, a main control unit and a target algorithm detection module, wherein the camera, the positioning module, the radar and the main control unit are mounted on a road cruise vehicle, and the target algorithm detection module is set on a user-operated PC; The camera is used to automatically take pictures of the road and transmit the pictures to the user's PC through the local area network for algorithm analysis to determine whether there are cracks or potholes. When the target algorithm detection module detects road damage in the image taken by the camera, it reads the positioning module data and sends the information to the EMQX server built on the Tencent Cloud lightweight application server through the ESP chip and further uploads it to the user interface, while displaying the image and detection results to the user.

2. The road self-inspection system based on YOLOv9 according to claim 1 is characterized in that: The main control unit is a Luban Cat microcomputer equipped with a Linux system and capable of visualizing the state of the cruise vehicle and remotely controlling the cruise vehicle through a local area network and a virtual machine.

3. The road self-inspection system based on YOLOv9 according to claim 1, characterized in that: The sensors used in the positioning module are odom, GPS and IMU.

4. The road self-inspection system based on YOLOv9 according to claim 1, characterized in that: The ESP chip packages the collected data and sends it to the cloud EMQX server, performs data analysis in the cloud, and displays real-time working information on the PC.

5. The road self-inspection system based on YOLOv9 according to claim 1, characterized in that: The target algorithm detection module adopts the YOLOv9 target detection algorithm, integrates the programmable gradient information PGI and the generative potential embedding GLEAN architecture for object detection into YOLOv9, further enhances its performance in object detection tasks, and forms the programmable gradient information PGI and its network architecture method; The programmable gradient information PGI and its network architecture method include the following four different network designs: (1) PAN: used to improve feature fusion to improve the performance of object detection; (2) RevCol: A network design used to reduce information loss, which maintains information flow without loss through a reversible column structure; (3) Deep supervision: improving learning efficiency and final model performance by inserting additional supervision signals at multiple layers of the network; (4) Programmable Gradient Information (PGI): used for forward propagation without skip links.

6. The road self-inspection system based on YOLOv9 according to claim 5, characterized in that: The programmable gradient information PGI includes a main branch, an auxiliary reversible branch and multi-level auxiliary information; Wherein, the main branch is used for the architecture of reasoning; The auxiliary reversible branch is used to generate reliable gradients to supply the main branch for back propagation; The multi-level auxiliary information is used to control the main branch to learn the plannable multi-level semantic information.

7. The road self-inspection system based on YOLOv9 according to claim 5, characterized in that: The YOLOv9 target detection algorithm includes detection model, training model, model evaluation, model lightweight, model deployment and detection effect; The detection models include three models: YOLOv9s, YOLOv9m, and YOLOv9 l. The difference between them is the different parameter amounts. YOLOv9s is the model with the smallest parameter amount. The depth and width of the three models are successively deepened, and the accuracy is also successively improved; The model lightweighting is achieved by quantizing the model and compressing the model; The model deployment is used to apply the trained model to practical applications, including task encapsulation and integration, platform adaptation and optimization, to provide users with efficient and accurate image recognition and detection services.

8. The road self-inspection system based on YOLOv9 according to claim 7, characterized in that: The YOLOv9s network structure consists of four parts, including the input end, the backbone feature extraction network Backbone, the bottleneck layer Neck and the detection module output layer Head.

9. The road self-inspection system based on YOLOv9 according to claim 8, characterized in that: The backbone feature extraction network Backbone includes a Focus module, a convolution module Conv, a C3 module and an SPP module; The Focus module is used to slice the image; The convolution module Conv module integrates three operations: convolution, normalization and activation function, and the activation function is the ReLU function; The C3 module includes three convolutions and several bottleneck modules. The bottleneck modules use 1×1 and 3×3 convolutions to extract features and use residual connections to enhance the network feature extraction capability. The SPP module is also called spatial pyramid pooling, which improves the perception field of the feature map by performing pooling operations on the feature map through pooling kernels of different sizes and then concat.

10. The road self-inspection system based on YOLOv9 according to claim 8, characterized in that: The bottleneck layer Neck adopts a bottom-up connection structure of the feature pyramid. The output layer Head of the detection module includes several detection heads, which predict feature maps of different sizes and types, corresponding to large targets, medium targets and small targets. The large-size feature maps are used to predict small targets, and the small-size feature maps are used to predict large targets. Each grid of the feature map corresponds to 3 anchor sizes with different width-to-height ratios. At the same time, prediction and regression are performed in the channel dimension of the feature map through the target position and classification information of the anchor prior box.

Images