A YOLO-based intelligent target labeling method and system

Abstract

The application relates to the technical field of computer vision, in particular to a YOLO-based intelligent target labeling method and system. The method comprises the following steps: obtaining image data of a to-be-labeled image; generating a first labeling box based on a labeling operation of a user on the image data, the first labeling box being associated with first label information; in response to a reuse instruction of the user on the first labeling box, obtaining second label information specified by the user, and generating at least one second labeling box with the same number of the second label information based on the first labeling box, and respectively associating each second label information to the corresponding second labeling box; obtaining labeling data based on the first labeling box, the second labeling box, the first label information and the second label information, the labeling data being used for training a YOLO target detection model. The labeling box reuse mechanism reduces the repeated labeling operation of the user, and effectively improves the labeling efficiency.

Description

Technical Field

[0001] This application relates to the field of computer vision target annotation technology, and in particular to a YOLO-based intelligent target annotation method and system. Background Technology

[0002] With the rapid development of computer vision technology, the YOLO object detection model has been widely used in various scenarios. As the core foundation for training the YOLO object detection model, the annotation efficiency and accuracy of the data directly affect the training effect and detection accuracy of the YOLO object detection model. Therefore, intelligent object annotation methods based on YOLO have become one of the research focuses in the field of computer vision.

[0003] In existing technologies, YOLO-based intelligent object annotation methods typically require users to manually generate bounding boxes on the image data of the image to be annotated and associate label information with the bounding boxes. When it is necessary to annotate targets with the same geometric parameters, users need to repeat the annotation operation to generate new bounding boxes and associate label information in order to obtain the annotation data used to train the YOLO object detection model.

[0004] However, in existing technologies, when multiple targets with the same geometric parameters need to be labeled, repeated labeling operations consume a lot of time and manpower, reducing labeling efficiency. At the same time, the lack of identification and differentiation between the generated multiple bounding boxes can easily lead to confusion of the bounding boxes, affecting the accuracy of the labeled data and thus affecting the training effect of the YOLO object detection model, failing to meet the requirements of efficient and accurate labeling. Summary of the Invention

[0005] This application provides a YOLO-based intelligent target annotation method and system to solve the technical problems of repetitive operations and low annotation efficiency caused by the lack of a bounding box reuse mechanism in the existing target annotation process.

[0006] To achieve the above objectives, in a first aspect, this application provides a YOLO-based intelligent target annotation method, comprising: In response to the user's selection of an image to be labeled, the image data of the image to be labeled is obtained; Based on the user's annotation operations on the image data, a first annotation box is generated, and the first annotation box is associated with first label information; In response to the user's instruction to reuse the first annotation box, perform the following operations: Obtain the second label information specified by the user, and generate at least one second label box based on the first label box. The number of second label boxes is the same as the number of second label information, and each second label information is associated with the corresponding second label box. Based on the first bounding box, the second bounding box, the first label information, and the second label information, labeled data is obtained, and the labeled data is used to train the YOLO object detection model.

[0007] Preferably, generating a first bounding box based on user annotation operations on image data includes: Obtain the annotation type selected by the user, which may be a rectangle or a polygon; Obtain geometric information provided by the user through annotation operations. When the annotation type is rectangle, the geometric information includes position parameters and size parameters. When the annotation type is polygon, the geometric information includes vertex coordinate sequence. The first annotation box is generated based on the annotation type and geometric information.

[0008] Preferably, when the annotation type is polygon, generating a first annotation box includes: Obtain the vertex coordinate sequence defined by the user on the image data; Based on the vertex coordinate sequence, a closed polygonal region is formed to generate the first bounding box.

[0009] Preferably, generating at least one second annotation box based on the first annotation box includes: Copy the geometric information of the first annotation box; Based on the copied geometric information, at least one second annotation box is generated.

[0010] Preferably, before associating each second label information to its corresponding second annotation box, the method further includes: The second annotation box is automatically selected, and a prompt message is output to remind the user to specify label information for the second annotation box.

[0011] Preferably, the annotation data is obtained based on the first annotation box, the second annotation box, the first label information, and the second label information, including: The first annotation box, the second annotation box, the first label information, and the second label information are serialized to generate serialized data; The serialized data is saved to a storage medium to obtain the labeled data.

[0012] Preferably, it further includes: Based on different save trigger events, the current annotation data is stored as the corresponding historical annotation version. Save trigger events include manual saving by the user and automatic saving by the system.

[0013] Preferably, it further includes: In response to the user's version restoration command, the current annotation data is restored to the annotation data corresponding to the specified historical annotation version.

[0014] Preferably, when saving the serialized data to the storage medium fails, the step of saving the serialized data to the storage medium is re-executed, and a message indicating that the saving failed is output.

[0015] Secondly, this application also provides a YOLO-based intelligent target annotation system, applied to the YOLO-based intelligent target annotation method of the first aspect, including: an annotation module; The annotation module is configured as follows: In response to the user's selection of an image to be labeled, the image data of the image to be labeled is obtained; Based on the user's annotation operation, a first annotation box is generated on the image data, and the first annotation box is associated with the first label information; In response to the user's instruction to reuse the first annotation box, perform the following operations: Obtain the second label information specified by the user, and generate at least one second label box based on the first label box. The number of second label boxes is the same as the number of second label information, and each second label information is associated with the corresponding second label box. Based on the first bounding box, the second bounding box, the first label information, and the second label information, labeled data is obtained, and the labeled data is used to train the YOLO object detection model.

[0016] As can be seen from the above technical solutions, this application provides a YOLO-based intelligent object annotation method and system. This method reuses annotation boxes by reusing instructions. Users can specify multiple second label information at once, and the system automatically generates a corresponding number of uniquely identified second annotation boxes, completing a one-to-one association between labels and annotation boxes. This method reduces repetitive interactive operations, significantly reduces time and labor costs, improves annotation efficiency and data accuracy in multi-object scenarios, and thus enhances the training effect of the YOLO object detection model, meeting the requirements for efficient and accurate annotation. Attached Figure Description

[0017] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart of the YOLO-based intelligent target annotation method provided in this application; Figure 2 A breakdown flowchart of step S200 in the method provided in the embodiments of this application; Figure 3 A breakdown flowchart of step S300 in the method provided in the embodiments of this application; Figure 4 This is a breakdown flowchart of step S400 in the method provided in the embodiments of this application. Detailed Implementation

[0019] The embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described below do not represent all embodiments consistent with this application. They are merely examples of systems and methods consistent with some aspects of this application.

[0020] It should be noted that the brief descriptions of terms in this application are only for the convenience of understanding the embodiments described below, and are not intended to limit the embodiments of this application. Unless otherwise stated, these terms should be understood in their ordinary and common meaning.

[0021] The terms “comprising” and “having”, and any variations thereof, are intended to cover but not exclude inclusion, for example, a product or device that includes a range of components is not necessarily limited to all of the components that are clearly listed, but may include other components that are not clearly listed or that are inherent to such product or device.

[0022] In the field of computer vision object annotation technology, existing annotation methods used to train YOLO object detection models often suffer from low annotation efficiency and poor accuracy. These methods typically require users to manually generate bounding boxes on the image data of the image to be annotated and associate label information with these boxes. When annotating multiple targets with the same geometric parameters, it is difficult to guarantee both efficiency and accuracy. When multiple targets with identical geometric parameters need to be annotated, users must repeatedly perform the annotation operation to generate new bounding boxes and associate label information, consuming significant time and manpower. Furthermore, the lack of differentiation between the generated bounding boxes easily leads to confusion, affecting the accuracy of the annotation data. This results in a significant investment of manpower in repetitive work for on-site annotation, increasing both the manpower and time costs of annotation and delaying the training progress of the YOLO object detection model when handling large batches of annotation tasks due to low efficiency, thus affecting the continuity and stability of model training.

[0023] To solve the above problems, see [link to relevant documentation]. Figure 1 This application provides a YOLO-based intelligent target annotation method in some embodiments, including: S100: In response to the user's selection operation of the image to be labeled, obtain the image data of the image to be labeled.

[0024] Specifically, after a user launches the annotation system, the annotation interface, implemented using Vue.js, displays a file management panel and a Canvas in the browser. Users can perform selection operations on the images to be annotated through the file management panel, such as clicking the upload button to select an image from the local disk, or dragging and dropping image files directly to a specified area. The system supports multiple data formats, including JPG, PNG, and BMP static image formats, as well as MP4 and AVI video formats. When a user selects a video file, the system extracts the video frames as a sequence of image frames to be annotated. Users can select keyframes frame by frame for annotation, thereby expanding the system's application scenarios and meeting the annotation needs of video data. After capturing the user's selection command, the system reads the selected image file or video frame and loads it as image data. The system then renders this image data onto the Canvas drawing function for display. After completing the above operations, the system successfully obtains the image data of the image to be annotated.

[0025] In some embodiments, the system provides a label management function. Users access the label management page and add labels for target categories. The system receives the label name and color entered by the user, generates a unique label identifier for the label, and saves the label configuration to storage. After completing the label configuration, users can see the created labels in the label list panel during subsequent labeling processes.

[0026] S200. Based on the user's annotation operation on the image data, a first annotation box is generated, and the first annotation box is associated with first label information.

[0027] Specifically, when a user performs annotation operations on image data, the system responds by generating a first annotation box based on the boundaries of the area drawn by the user on the canvas, using Canvas technology. While the user is dragging the mouse to draw, the system utilizes Canvas's real-time drawing capabilities to create a preview box on the canvas. The border color and fill transparency of the preview box dynamically change with mouse movement, providing immediate visual feedback. After the user releases the mouse, the system officially generates the first annotation box, displaying it overlaid on the image data with a highlighted border and semi-transparent fill. The user can adjust the size and position of the annotation box by dragging its boundaries or top corners. During this adjustment, the system continuously updates the preview effect of the annotation box to ensure the user can accurately select the target object. Furthermore, the system provides grid lines and guide lines on the Canvas to assist user positioning and supports keyboard shortcuts (such as the Delete key to delete the selected annotation box and Ctrl+Z to undo the previous operation) to improve ease of use. After the first annotation box is generated, the system displays a label list panel. The user selects a category from the label list panel, and the system obtains the selected category as the first label information and establishes a relationship between the first label information and the first annotation box.

[0028] In some embodiments, see Figure 2 Based on the user's annotation operations on the image data, a first bounding box is generated, including: S201. Obtain the annotation type selected by the user. The annotation type includes rectangle or polygon.

[0029] Specifically, users specify the annotation type by clicking the rectangle tool button or polygon tool button on the interface toolbar. After detecting the user's click, the system sets the current annotation type to the selected type. For example, if the user clicks the rectangle tool button, the system sets the current annotation type to rectangle; if the user clicks the polygon tool button, the system sets the current annotation type to polygon.

[0030] S202. Obtain the geometric information provided by the user through annotation operations.

[0031] When the annotation type is rectangle, the geometric information includes position parameters and size parameters; when the annotation type is polygon, the geometric information includes vertex coordinate sequence.

[0032] Specifically, when the annotation type is rectangle, the user moves the mouse pointer to the vicinity of the top-left corner of the target object, presses the left mouse button, and drags it to the bottom-right corner. The system records the coordinates of the mouse press point and the current mouse point coordinates in real time, and draws a rectangular preview box on the canvas. After the user releases the left mouse button, the system calculates the x-coordinate, y-coordinate, width, and height of the top-left corner of the rectangle based on the coordinates of the mouse press and release points, using these positional and dimensional parameters as geometric information. When the annotation type is polygon, the user moves the mouse pointer to the outline boundary of the target object and clicks multiple positions in sequence. The system records the coordinates of each click point and connects adjacent click points with line segments to form a polyline. After the user completes clicking all vertices, double-clicking the mouse or clicking the "Done" button, the system uses the recorded vertex coordinate sequence as geometric information.

[0033] S203. Generate the first annotation box based on the annotation type and geometric information.

[0034] Specifically, the system generates a first annotation box based on the annotation type obtained in step S201 and the geometric information obtained in step S202. If the annotation type is a rectangle, the system generates a rectangular first annotation box based on the position parameters and size parameters; if the annotation type is a polygon, the system generates a polygonal first annotation box based on the vertex coordinate sequence.

[0035] Regardless of whether the first bounding box is a rectangle or a polygon, the system will overlay it onto the image data with a highlighted border and semi-transparent fill. Users can fine-tune the bounding box by dragging its boundaries or vertices. After the first bounding box is generated, the system displays a label list panel. Users select a category from the label list panel, and the system associates the corresponding label information of that category with the first bounding box, completing the creation and label association of the first bounding box.

[0036] In some embodiments, when the annotation type is polygon, generating a first annotation box includes: Obtain the sequence of vertex coordinates defined by the user on the image data.

[0037] Based on the vertex coordinate sequence, a closed polygonal region is formed to generate the first bounding box.

[0038] Specifically, in polygon annotation mode, users click on multiple locations sequentially on the image data using the mouse or touch. The system records the coordinates of each click on the image data, with each coordinate corresponding to a vertex of the polygon. After defining all vertices, the user performs a completion operation, such as double-clicking the left mouse button or clicking the "complete" button on the interface. Upon detecting the completion operation, the system connects the coordinates of the first and last vertices and connects adjacent vertices in the order they were clicked, forming a closed polygonal region. The system then generates the first polygonal bounding box based on this closed polygonal region. The first bounding box is displayed superimposed on the image data as a polygonal graphic, with its boundary lines connecting the defined vertices end to end, enclosing the area that is the outline of the target object specified by the user. After generating the first bounding box, the user can associate label information with this polygonal bounding box.

[0039] This embodiment supports polygon annotation types, which can adapt to target objects with irregular shapes and contours, making up for the shortcomings of rectangular annotations in accurately fitting complex boundaries, thereby expanding the applicability of the annotation system and improving the accuracy of annotation data.

[0040] S300, In response to the user's instruction to reuse the first annotation box, perform the following operations: Obtain the second label information specified by the user, and generate at least one second label box based on the first label box. The number of second label boxes is the same as the number of second label information, and each second label information is associated with the corresponding second label box.

[0041] Specifically, the user executes a reuse command on the first annotation box. This command can be triggered by clicking the reuse button on the toolbar, pressing a preset shortcut key, or right-clicking the first annotation box and selecting the copy menu. In response to this command, the system displays a tag list panel for the user to select tags. After the user selects one or more tags, the system retrieves the second tag information specified by the user. Based on the first annotation box, the system generates the same number of second annotation boxes as the number of second tag information items. Each second annotation box does not contain tag information, and the system assigns each second annotation box a unique identifier different from the first and other second annotation boxes. The system associates each second tag information selected by the user with a corresponding second annotation box. After completing the above operations, the system generates multiple second annotation boxes with associated tags at once.

[0042] In some embodiments, see Figure 3 Generate at least one second annotation box based on the first annotation box, including: S301. Copy the geometric information of the first annotation box.

[0043] Specifically, after detecting the reuse instruction, the system reads the geometric information of the first bounding box. If the first bounding box is a rectangle, the geometric information includes the ordinate of its upper left corner, the x-coordinate of its upper left corner, its width, and its height; if the first bounding box is a polygon, the geometric information includes its vertex coordinate sequence.

[0044] S302. Based on the copied geometric information, generate at least one second annotation box.

[0045] Specifically, the system generates the same number of second bounding boxes as the user-specified number of second label information, based on the geometric information copied in step S301. The geometric information of each second bounding box is identical to that of the first bounding box: if the first bounding box is a rectangle, each second bounding box has the same top-left ordinate, top-left abscissa, width, and height as the first bounding box; if the first bounding box is a polygon, each second bounding box has the same vertex coordinate sequence as the first bounding box. During generation, the system assigns a unique identifier to each second bounding box, which is different from both the first and other second bounding boxes, and initializes the label information of each second bounding box to empty. After generation, the system adds all second bounding boxes to the bounding box list of the current image and draws them on the canvas. These steps ensure that the number of second bounding boxes matches the number of second label information specified by the user. Subsequently, the system associates each user-specified second label information with its corresponding second bounding box.

[0046] In some embodiments, before associating each second label information to its corresponding second annotation box, the method further includes: Automatically select each second label box and output a prompt message to remind the user to specify label information for each second label box.

[0047] Specifically, after generating each second annotation box but before associating it with second label information, the system automatically selects all second annotation boxes. Each selected second annotation box is highlighted with a high-brightness border. Simultaneously, the system displays a prompt message on the interface, such as "Please assign label information to the selected annotation boxes in sequence," guiding the user to perform the label selection operation. After seeing the prompt message, the user can assign labels to each second annotation box individually through the label list panel. This automatic selection and prompt mechanism reduces the steps required for users to manually click on each box, significantly improving the efficiency of batch annotation operations.

[0048] S400. Based on the first annotation box, the second annotation box, the first label information, and the second label information, the annotation data is obtained.

[0049] The labeled data is used to train the YOLO object detection model.

[0050] Specifically, the system integrates the first bounding box, all generated second bounding boxes, first label information, and the second label information corresponding to each second bounding box into annotation data. The annotation data is in a structured data format, such as JSON. This annotation data contains the correspondence between the geometric information of the first bounding box and its first label information, as well as the correspondence between the geometric information of each second bounding box and its associated second label information. This annotation data can be directly used by the YOLO object detection model training program.

[0051] In some embodiments, see Figure 4 Based on the first annotation box, the second annotation box, the first label information, and the second label information, the annotation data is obtained, including: S401. Serialize the first annotation box, the second annotation box, the first label information, and the second label information to generate serialized data.

[0052] Specifically, the system serializes the geometric information of the first annotation box, the first label information, the geometric information of each second annotation box and its associated second label information according to the agreed data format, converts the annotation objects into string form, and generates serialized data.

[0053] S402. Save the serialized data to the storage medium to obtain the labeled data.

[0054] Specifically, the system writes the generated serialized data into a storage medium. After saving, the serialized data becomes persistent labeled data, which, together with the original image data, constitutes training samples that can be used by the YOLO object detection model training program.

[0055] In some embodiments, the current annotation data is stored as a corresponding historical annotation version based on different save trigger events, including manual saving by the user and automatic saving by the system.

[0056] Specifically, the system identifies the trigger source of the current save operation. If the user actively clicks the save button or presses the save shortcut key, it is determined to be a user-manual save event. If the system detects changes in the annotation data and meets the preset automatic save conditions, it is determined to be a system-automatic save event. The system adopts a real-time automatic save strategy, that is, each annotation operation (generating annotation boxes, modifying labels, reusing annotation boxes, adjusting box positions, etc.) is triggered immediately after completion, without waiting for a fixed time interval. This mechanism effectively avoids the loss of annotation data due to abnormal situations such as unexpected program closure, system crashes, or power outages, significantly improving data security. Based on different trigger events, the system stores the current annotation data as an independent historical annotation version, and each historical annotation version carries a corresponding timestamp and trigger type identifier.

[0057] In some embodiments, the system supports exporting labeled data in the format required for model training. Users select the labeled dataset through the dataset management interface and perform the export operation. The system packages all image data and their corresponding labeled data, organizes them according to the directory structure required by the YOLO model, and generates the dataset folder required for training. After exporting, users can directly use this dataset folder for training the YOLO object detection model.

[0058] In some embodiments, the system also provides model training task management functionality. Users create training tasks, set training parameters (such as number of iterations, batch size, learning rate, etc.), and select exported training and validation data. The system calls the backend training service to start the training process, monitors the training progress in real time, and displays the loss curve. After training is complete, users can evaluate the model's performance metrics on the validation set and iteratively optimize the labeled data or adjust the training parameters based on the evaluation results.

[0059] In some embodiments, in response to a user's version restoration command, the current annotation data is restored to the annotation data corresponding to a specified historical annotation version.

[0060] Specifically, the user selects a version from the list of historical annotation versions and executes the version restoration command. After receiving the command, the system obtains the annotation data corresponding to the historical annotation version specified by the user. The system replaces the current annotation data with the annotation data of the historical annotation version and redraws all the annotation boxes on the image canvas based on the replaced annotation data. After the replacement is completed, the current annotation state is restored to the state of the selected historical annotation version.

[0061] In some embodiments, when saving serialized data to the storage medium fails, the step of saving serialized data to the storage medium is re-executed, and a message indicating that the saving failed is output.

[0062] Specifically, when performing the serialized data saving operation, the system monitors the return status during the saving process. If the saving operation fails due to network interruption, insufficient storage space, or permission error, the system captures the failure event and automatically re-executes the step of saving the serialized data to the storage medium. If the saving is successful after the retry, the subsequent process continues. If the saving still fails after the retry, the system outputs a saving failure message on the user interface to inform the user of the current saving status.

[0063] In some embodiments, when a user needs to save the current state of the labeled dataset, they can perform the operation of creating a dataset version through the dataset management interface. The system responds to this operation by displaying a version creation dialog box. The user enters the version name and description information in the dialog box and confirms the creation. After receiving the version name and description information entered by the user, the system retrieves all image data and their corresponding annotation data in the current dataset and generates a new dataset version based on this data. The system assigns a unique version identifier to this version and records the version creation time. The newly created dataset version contains complete image files and annotation data snapshots of the current dataset. Through version management, users can trace the historical state of each major modification, support comparing annotation differences between different versions in large-scale datasets, and quickly revert to any historical version when needed, solving the problems of version chaos and difficulty in collaboration in traditional annotation methods. After version creation is completed, users can select this version to view or edit during subsequent annotation processes through the version switching function, in order to trace the historical annotation state or restore the dataset content to a specific point in time.

[0064] In some embodiments, this application provides a YOLO-based intelligent target annotation system, applied to the YOLO-based intelligent target annotation method provided in the above embodiments, including: In response to the user's selection of an image to be labeled, the image data of the image to be labeled is obtained; Based on the user's annotation operation, a first annotation box is generated on the image data, and the first annotation box is associated with the first label information; In response to the user's instruction to reuse the first annotation box, perform the following operations: Obtain the second label information specified by the user, and generate at least one second label box based on the first label box. The number of second label boxes is the same as the number of second label information, and each second label information is associated with the corresponding second label box. Based on the first bounding box, the second bounding box, the first label information, and the second label information, labeled data is obtained, and the labeled data is used to train the YOLO object detection model.

[0065] Specifically, the annotation module is implemented based on the Vue.js framework, displaying a file management panel and a Canvas in the browser. Users select images to be annotated through the file management panel, and the annotation module loads the selected images and renders them on the Canvas. When a user performs an annotation operation, the annotation module generates a first annotation box based on the boundary of the area drawn by the user on the canvas, and displays a label list panel for the user to select a category. The selected category is then used as the first label information and associated with the first annotation box.

[0066] When a user executes a reuse command on the first annotation box, the annotation module retrieves one or more second label information specified by the user and generates the same number of second annotation boxes as the first annotation box. Each second annotation box copies the geometric information of the first annotation box, does not contain label information, and is assigned a unique identifier different from the first and other second annotation boxes. The annotation module automatically selects all second annotation boxes and outputs prompts to guide the user to specify label information for each second annotation box, retrieves the user-specified second label information, and associates it with the corresponding second annotation box.

[0067] Finally, the annotation module integrates the first bounding box, all second bounding boxes, the first label information, and the information of each second label into structured annotation data. This annotation data is stored in JSON format. After exporting to generate a dataset folder, it can be directly used for training the YOLO object detection model. Through the above configuration, this system achieves efficient and accurate intelligent object annotation.

[0068] The above specific embodiments further illustrate the purpose, technical solution and beneficial effects of this application. It should be understood that the above are only specific embodiments of this application and are not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of this application should be included within the scope of protection of this application.

Claims

1. A YOLO-based intelligent target annotation method, characterized in that, include: In response to the user's selection of an image to be labeled, the image data of the image to be labeled is obtained; Based on the user's annotation operation on the image data, a first annotation box is generated, and the first annotation box is associated with first label information; In response to the user's instruction to reuse the first annotation box, perform the following operations: Obtain the second label information specified by the user, and generate at least one second label box based on the first label box. The number of second label boxes is the same as the number of second label information, and each piece of second label information is associated with the corresponding second label box. Based on the first bounding box, the second bounding box, the first label information, and the second label information, labeled data is obtained, and the labeled data is used to train the YOLO object detection model.

2. The intelligent target annotation method based on YOLO according to claim 1, characterized in that, The step of generating a first bounding box based on the user's annotation operations on the image data includes: Obtain the annotation type selected by the user, where the annotation type includes rectangle or polygon; Obtain geometric information provided by the user through annotation operations, wherein when the annotation type is rectangle, the geometric information includes position parameters and size parameters, and when the annotation type is polygon, the geometric information includes vertex coordinate sequence; The first annotation box is generated based on the annotation type and the geometric information.

3. The intelligent target annotation method based on YOLO according to claim 2, characterized in that, When the annotation type is polygon, generating the first annotation box includes: Obtain the vertex coordinate sequence defined by the user on the image data; Based on the vertex coordinate sequence, a closed polygonal region is formed to generate the first annotation box.

4. The intelligent target annotation method based on YOLO according to claim 2, characterized in that, The step of generating at least one second annotation box based on the first annotation box includes: Copy the geometric information of the first annotation box; Based on the copied geometric information, at least one second annotation box is generated.

5. The intelligent target annotation method based on YOLO according to claim 1, characterized in that, Before associating each of the second label information to its corresponding second annotation box, the method further includes: The second label box is automatically selected, and a prompt message is output, which prompts the user to specify label information for the second label box.

6. The intelligent target annotation method based on YOLO according to claim 1, characterized in that, The step of obtaining annotation data based on the first annotation box, the second annotation box, the first label information, and the second label information includes: The first annotation box, the second annotation box, the first label information, and the second label information are serialized to generate serialized data; The serialized data is saved to a storage medium to obtain the labeled data.

7. The intelligent target annotation method based on YOLO according to claim 6, characterized in that, Also includes: Based on different save trigger events, the currently labeled data is stored as the corresponding historical labeled version. The save trigger events include manual saving by the user and automatic saving by the system.

8. The intelligent target annotation method based on YOLO according to claim 7, characterized in that, Also includes: In response to the user's version restoration command, the current annotation data is restored to the annotation data corresponding to the specified historical annotation version.

9. The intelligent target annotation method based on YOLO according to claim 6, characterized in that, If saving the serialized data to the storage medium fails, the step of saving the serialized data to the storage medium is re-executed, and a message indicating that the saving failed is output.

10. A YOLO-based intelligent target annotation system, applied to the YOLO-based intelligent target annotation method according to any one of claims 1-9, characterized in that, Includes: annotation module; The annotation module is configured as follows: In response to the user's selection of an image to be labeled, the image data of the image to be labeled is obtained; Based on the user's annotation operation, a first annotation box is generated on the image data, and the first annotation box is associated with first label information; In response to the user's instruction to reuse the first annotation box, perform the following operations: Obtain the second label information specified by the user, and generate at least one second label box based on the first label box. The number of second label boxes is the same as the number of second label information, and each piece of second label information is associated with the corresponding second label box. Based on the first bounding box, the second bounding box, the first label information, and the second label information, labeled data is obtained, and the labeled data is used to train the YOLO object detection model.

Images