A picture AI recognition training marking system and method
By replacing traditional point-by-point annotation with contour graphics, and combining pixel ratio conversion and AI recognition verification, the problems of low efficiency and unstable accuracy of manual labeling methods are solved, and efficient and accurate AI training data preparation is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HANGZHOU HUIKAIZHANG TECHNOLOGY CO LTD
- Filing Date
- 2026-02-12
- Publication Date
- 2026-06-02
Smart Images
Figure CN122134844A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of artificial intelligence, in particular to a picture AI recognition training marking system and method. BACKGROUND
[0002] In the picture AI recognition model training process, accurate marking of sample pictures is the core prerequisite for determining the recognition accuracy of the model, and the accuracy, completeness and acquisition efficiency of the marking data directly affect the training effect and iteration period of the AI model. At present, manual point-by-point marking is the most widely used basic marking method in picture AI recognition training, especially suitable for industrial part detection, graphic target recognition and other fields that have basic requirements for marking accuracy, and is also the mainstream choice in most scenarios.
[0003] The core process of manual point-by-point marking is that the operator manually marks the key nodes of the target contour on the sample picture with the help of basic drawing tools, then connects the nodes to form a complete marking area, and manually records information such as target shape type and approximate position, finally completes the marking work of a single target. This method is still widely used in small and medium-sized sample marking scenarios due to its low operation threshold and the need for complex tool support, and can to some extent meet the marking needs of complex shape targets.
[0004] However, manual point-by-point marking has many inherent defects and cannot meet the actual needs of large-scale and high-precision AI training: first, the marking efficiency is extremely low. For a sample picture containing multiple targets, each target needs to be marked, connected and information recorded point by point, which is time-consuming and has a large amount of repetitive operations. When faced with the massive amount of samples required for AI training, the marking period is too long and the model iteration progress is slowed down; second, the marking accuracy is poor. Affected by human factors such as visual judgment, hand control ability and marking habit differences, the positioning deviation of key nodes is difficult to avoid, and the consistency of the same batch of sample data marked by different personnel is poor, even the marking results of the same personnel at different times also fluctuate, which directly reduces the stability and final recognition accuracy of the AI model training; third, the labor intensity is large. Long-term focus on point-by-point marking and information recording can easily lead to operator fatigue, further exacerbating marking deviation, forming a vicious cycle of "fatigue-error-rework"; fourth, data collation is tedious. After manual marking, the node coordinates, shape types and other data need to be manually collated, the format is chaotic and has no unified standard, and additional effort is needed to standardize the processing to adapt to the AI model training requirements, increasing the overall process cost.
[0005] In summary, the existing artificial marking method generally has the problems of low efficiency, unstable precision, high labor intensity, and complicated data arrangement, which has become the core bottleneck restricting the large-scale and efficient training of AI recognition models. Therefore, there is an urgent need for a marking system and method that can break through the limitations of artificial marking, taking into account the marking accuracy, efficiency, and data specification, to solve the problems of the existing technology. SUMMARY
[0006] In order to solve the problems of low efficiency, unstable precision, high labor intensity, and complicated data arrangement of the existing artificial marking method, the purpose of the present application is to provide a marking system and method for picture AI recognition training.
[0007] In order to achieve the above-mentioned purpose, the present application adopts the following technical solution: a marking system for picture AI recognition training, comprising a front-end interaction module and a server module; the front-end interaction module is provided with a menu area, a mark graphic candidate column and a canvas area, the menu area is configured with a picture selection key, a marking key, an upload key and an identification key, the mark graphic candidate column stores a plurality of candidate basic shape icons, each shape icon corresponds to a background transparent outline graphic, the outline graphic is used to replace the traditional point-by-point labeling method to realize the marking of the target features of the sample picture, the canvas area is used to display the sample picture and the outline graphic, and supports performing operations on the outline graphic through a mouse or a touch screen gesture and realizing 2D / 3D conversion of the outline graphic in real time through a transform attribute, the transform attribute can realize rotation, scaling, moving and tilting operations of the outline graphic; the server module is installed with a server program and a picture AI recognition program, the server program is used to receive the sample picture and the operation information based on the transform attribute transmitted by the front-end interaction module, and the picture AI recognition program is used to identify the sample picture and return the identification result; the front-end interaction module and the server module are in communication connection to realize the transmission of the sample picture and the transform attribute operation information.
[0008] Preferably, the front-end interaction module is further configured with a data processing unit, the data processing unit is used to obtain the scale coefficient sc of the actual pixels of the sample picture and the screen display pixels, and to perform pixel scale conversion processing on the related operation information of the outline graphic obtained based on the transform attribute.
[0009] Preferably, the front-end interaction module is further configured with a storage unit, the storage unit stores the marking data through a labs array, the marking data takes the operation information of the transform attribute as the core, and further includes the moving, scaling, rotating and tilting information converted by the scale coefficient sc, as well as the pixel width and height of the sample picture, the shape type and pixel size corresponding to the outline graphic.
[0010] A picture AI recognition marking method, comprising the following steps: S1, import sample images from folders or albums using the image selection key, display the sample images on the canvas area, and the data processing unit obtains the scaling factor sc. S2, click the target shape icon in the candidate graphic marker bar, and the corresponding outline graphic will be displayed in the upper center of the canvas area according to the preset size; S3 listens for screen touch or mouse operations and uses the transform property to instantly rotate, scale, move, or tilt the outline graphic, making the outer outline of the outline graphic completely overlap with the target shape in the sample image. S4, click the label button, multiply the original operation information in the transform property by the scaling factor sc and convert it into operation information in the pixel dimension of the sample image, and save it to the labs array; S5. Repeat steps S2-S4 to complete the labeling of multiple targets in the sample image. Click the upload button to upload the sample image and the labs array storing transform attribute operation information to the server module for training the image AI recognition program.
[0011] Preferably, in step S4, the movement information of the transform attribute includes the distance between the center of the outline graphic and the center of the canvas area. After multiplying the distance by the scaling factor sc, it is converted into the pixel distance between the outline graphic and the center of the sample image and saved to the labs array.
[0012] Preferably, in step S4, the scaling information of the transform attribute includes the screen display size of the outline graphic. After multiplying the display size by the scaling factor sc, it is converted into the corresponding pixel size of the outline graphic on the sample image and saved to the labs array.
[0013] Preferably, in step S2, the preset size of the outline graphic is a standard size that adapts to the display ratio of the canvas area, and the size of the outline graphic can be adjusted in real time by screen touch or mouse operation combined with the transform property.
[0014] A method for verifying the accuracy of image AI recognition utilizes the transform property to pass operational information about the contour graphic, and includes the following steps: T1: Import the sample image to be verified using the image selection key. The sample image is displayed on the canvas area, and the data processing unit obtains the scaling factor sc. T2, Click the recognition button to upload the sample image to the server module, but do not upload the corresponding labeling data; T3, after the server program detects that it has only received sample images, it forwards them to the image AI recognition program; T4, the image AI recognition program, after recognizing the sample image, returns the recognition result including the shape type; T5's front-end interaction module generates a corresponding contour graphic based on the recognition results. The transform property is used to instantly rotate, scale, move, or tilt the contour graphic to align it with the target shape in the sample image. Users can judge the recognition accuracy by observing the overlap between the contour graphic and the target shape.
[0015] Preferably, in step T5, the front-end interaction module performs pixel ratio conversion on the operation information of the transform attribute based on the ratio coefficient sc, and adjusts the size and position of the outline graphic so that the pixel ratio of the outline graphic matches that of the sample image.
[0016] Preferably, the recognition result also includes the approximate position and size information of the target in the sample image. The front-end interaction module initially locates the display position and size of the outline graphic based on the information, and then makes fine adjustments through the transform attribute to make the outline graphic accurately aligned with the target shape.
[0017] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: 1. This invention uses contour graphics to replace the traditional manual point-by-point annotation method. It eliminates the need for operators to manually mark a series of nodes of the target contour. Simply select a contour graphic that is similar to the target features of the sample image and make simple adjustments to complete the marking. This completely solves the problems of low efficiency, large amount of repetitive operation, and high labor intensity of traditional point-by-point annotation. It significantly shortens the marking cycle of a single sample image and avoids human positioning deviation caused by point-by-point marking. It improves the marking accuracy and data consistency and provides more stable annotation data support for AI recognition training. 2. This invention utilizes the transform property built into the web page object to transmit operation information of the outline graphic. This property can instantly realize transformation operations such as rotation, scaling, movement and tilting of the outline graphic. The operation response is smooth and there is no need to add a complicated graphic processing program, which simplifies the adjustment process of the outline graphic. At the same time, it can accurately capture and save all operation information in the transform property, ensuring the integrity and accuracy of the labeling data. After being converted by the scaling factor, it is uploaded to the server, which can directly adapt to the needs of the AI training program, reduce the data processing links, and further improve the overall efficiency of AI training. 3. This invention calculates the ratio coefficient sc between the actual pixels of the sample image and the pixels displayed on the screen through the front-end data processing unit, accurately converts the operation information in the transform property with sc, and combines the labs array to store multi-dimensional standardized labeling data, effectively avoiding labeling distortion caused by the deviation between the display ratio and the actual pixels, further ensuring that the labeling data is fully adapted to the pixel dimension of the sample image, and improving the training quality of AI models from the source. 4. This invention integrates labeling training and AI recognition verification functions. The front end triggers the server-side AI recognition program through the recognition key, and the contour graphic display is adapted by the scaling factor sc. Users can intuitively judge the recognition accuracy through the contour overlap, forming a complete closed loop of labeling, training and verification. No additional third-party tools are required, which improves the efficiency of the entire AI training process. Attached Figure Description
[0018] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Fig. 1 This is a schematic diagram of the user interface of the present invention.
[0019] Fig. 2 This is a logic diagram of the training process of the present invention.
[0020] Fig. 3 This is a logic diagram of the verification process of the present invention. Detailed Implementation
[0021] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification.
[0022] Please see Figs. 1 to 3 It should be understood that the structures, proportions, sizes, etc., illustrated in the accompanying drawings are merely for illustrative purposes to aid those skilled in the art and to facilitate understanding and reading. They are not intended to limit the scope of the invention and therefore have no substantial technical significance. Any modifications to the structure, changes in proportions, or adjustments to size, without affecting the effectiveness and purpose of the invention, should still fall within the scope of the technical content disclosed in this invention. Furthermore, the terms such as "upper," "lower," "left," "right," "middle," and "one" used in this specification are merely for clarity and not intended to limit the scope of the invention. Changes or adjustments to their relative relationships, without substantially altering the technical content, should also be considered within the scope of the invention's implementation.
[0023] This invention provides a technical solution: a labeling system for image AI recognition training operating environment configuration including front-end device, server device, and data storage. The front-end device can be a computer, compatible with Windows 10 and above and equipped with a mouse, or a tablet, compatible with Android 12 / iOS 16 and above and supporting touch screens, with a screen resolution of no less than 1920×1080 to ensure clear display of sample images and outline graphics in the canvas area; the front-end runtime environment supports HTML5 and JavaScript, and uses the CSS transform property to realize graphic interactive adjustments, and uses the Canvas API to render canvas content. The core utilizes the transform property to realize real-time transformations such as rotation, scaling, movement and skewing of outline graphics, replacing the traditional point-by-point annotation method.
[0024] Server equipment: Cloud server, configured with Intel Xeon E5-2670v3 CPU, 16GB memory, 1TB hard drive, CentOS 8.0 operating system, server program developed using Java Spring Boot framework, image AI recognition program built on YOLOv8 object detection model, supports recognition of common basic shapes such as rectangle, circle, triangle, etc. Data transmission between server and front end is achieved through HTTP / HTTPS protocol, with a transmission rate of not less than 1Mbps.
[0025] Data storage: The front-end temporary storage unit uses the browser's local storage, and the labeled data is stored through the labs array. The core of the array stores the operation information related to the transform property, replacing the traditional feature point coordinate data labeled point by point. The server uses a MySQL database to store sample images and corresponding labeled data in association. The sample images are in JPG / PNG format, with a pixel range of 500×500 to 4000×4000, which is convenient for AI training.
[0026] An image AI recognition and labeling system consists of a front-end interaction module and a server module. The front-end interaction module provides users with a visual interface, integrating operation entry, visual interaction, data processing, and temporary storage functions. The front-end interaction module is configured with a menu area, a marked graphic candidate bar, a canvas area, a data processing unit, and a storage unit. The menu area is located at the top of the interface and contains four function buttons, each with a clear icon and text prompt: the image selection button corresponds to the folder icon, which triggers a local file selection window when clicked, supporting batch import of sample images from folders or albums; the label button corresponds to the tag icon, which triggers a command to save the label data, primarily saving the operation information of the transform attribute; the upload button corresponds to the upload icon, which packages and uploads the sample images and the labs array data to the server when clicked; and the recognition button corresponds to the magnifying glass icon, which uploads only the sample images for AI recognition and verification when clicked.
[0027] The candidate shape bar is located on the left side of the interface and stores six commonly used basic shape icons: rectangle, circle, triangle, ellipse, rhombus, and polygon. Each icon is 48×48 pixels in size. Clicking it generates a corresponding transparent outline shape, which is used to replace the traditional point-by-point annotation of feature points. The line color is red and the line width is 2px. The initial preset size of the outline shape is the standard size adapted to the display ratio of the canvas area, that is, 1 / 10 of the width of the canvas area, so that it can be quickly adjusted to overlap with the target outline of the sample through the transform property.
[0028] The canvas area is located in the core central area of the interface. Its size adapts to the front-end device screen, with a maximum display range of 1920×1080 pixels. It supports adaptive scaling of sample images while maintaining the image aspect ratio. It also supports screen touch or mouse interaction. When using touch, a single finger drag moves the outline graphic, two fingers pinch / spread to zoom, and two fingers rotate to adjust the angle. When using the mouse, the left mouse button drags to move, the scroll wheel zooms, and the right mouse button is held down to rotate. All adjustment operations are responded to instantly through the transform property. During the operation, the position and size changes of the outline graphic are fed back in real time until the outer contour of the outline graphic completely overlaps with the target shape of the sample image, completing the annotation.
[0029] The data processing unit implements its functions using JavaScript. After importing sample images, it automatically reads the actual pixel width and height from the image metadata, denoted as W and H. At the same time, it obtains the pixel width and height of the image displayed on the canvas area, denoted as W' and H'. It calculates the scaling factor sc = W / W' = H / H' to ensure consistent width and height ratios and avoid stretching. Subsequently, the movement, scaling, and rotation information of the contour graphic captured by the transform property are multiplied by sc and converted into data of the actual pixel dimensions of the corresponding sample image, ensuring the accuracy of the labeling data.
[0030] The storage unit stores all the marking data of a single sample image through the labs array. Each element in the array corresponds to the marking information of a target. The core includes the transform attribute operation information after being converted by sc, including the movement coordinates, scaling ratio, rotation angle information, as well as the actual pixel width and height of the sample image, the shape type code corresponding to the outline graphic, and the actual pixel size of the outline graphic. The movement coordinates are based on the upper left corner of the sample image as the origin, the rotation angle is in degrees, and the shape type code is 1 for rectangles, 2 for circles, etc.
[0031] The server module integrates data receiving and forwarding, AI recognition and calculation, and data association and storage functions. The server module is configured with a server program, an image AI recognition program, and a data association and storage unit. The server program is developed using the Java Spring Boot framework and has functions for receiving, parsing, forwarding, and sending back data. After receiving the data uploaded by the front end, it first determines whether it contains the labs array. If it does, it associates the sample image with the labeled data containing transform attribute operation information and stores it in the MySQL database to form an AI training dataset. If it only contains sample images, it converts the image data to RGB channel format, normalizes it to 640×640 pixels, and then forwards it to the image AI recognition program.
[0032] The image AI recognition program is built on the YOLOv8 object detection model. After initial training, the model has the ability to recognize basic shape objects. After receiving sample images forwarded by the server program, it identifies the objects in the images through forward inference and generates recognition results. The results include the object shape type, position coordinates, and size. The position coordinates are based on the upper left corner of the sample image, and the size is the width and height. The recognition results are then formatted and sent back to the front end through the server program.
[0033] The data association storage unit uses a MySQL database. The data table contains fields such as sample image ID, image storage path, JSON string of labeled data, upload time, and recognition result. The recognition result is an optional storage item. The core of the JSON string of labeled data contains operation information related to the transform attribute. It supports data retrieval by shape type and upload time, which is convenient for batch calling of datasets during AI training.
[0034] Specific steps for implementing the marking method: The configuration of the image AI recognition training and labeling system, and the specific process of the image AI recognition training and labeling method are as follows: S1: Sample Image Import and Scale Factor Calculation. The user clicks the image selection button in the front-end menu area to select a 1920×1080 pixel industrial part sample image from a local folder. This image contains three rectangular part targets, which are automatically displayed in the canvas area. The canvas display size is 960×540 pixels. The front-end data processing unit automatically calculates the scale factor sc=1920 / 960=2, providing the basis for pixel conversion of subsequent transform attribute operation information.
[0035] S2: Outline Graphic Call and Display. When the user clicks the rectangle icon in the left-hand marker graphic candidate bar (corresponding shape type code 1), the system generates a red transparent outline graphic to replace the traditional point-by-point annotation of feature points. It is displayed in the upper center of the canvas area at a preset size, which is 1 / 10 of the canvas area width, i.e., 96 pixels wide. The outline graphic has clear lines and does not obscure the core features of the sample image. This preset size can be adjusted in real time through interactive operation.
[0036] S3: Outline Graphic Adjustment and Alignment. The user drags the outline graphic with the left mouse button to move it to the position of the first rectangular part in the sample image, zooms the outline graphic with the mouse wheel, and rotates it by holding down the right mouse button. In this embodiment, the part is upright and the rotation angle is 0°. All adjustment operations are responded to instantly by the transform property to achieve rapid transformation of the outline graphic until the outer contour of the outline graphic completely overlaps with the shape of the rectangular part. Release the mouse to complete the adjustment.
[0037] S4: Marking Data Saving. When the user clicks the marking button in the menu area, the data processing unit reads the movement distance, scaling ratio, and rotation angle from the transform property. The movement distance corresponds to the canvas display dimensions of X=200 pixels and Y=150 pixels, the scaling ratio is 1.2, and the rotation angle is 0°. These are multiplied by the scaling factor sc=2 to convert them into the actual pixel dimension data of the sample image, i.e., X=400, Y=300, scaling ratio=1.2, and rotation angle=0°. The movement information is converted into the pixel distance of the outline graphic relative to the center of the sample image, and the scaling information is converted into the corresponding pixel size of the outline graphic on the sample image. This data, along with the sample image width and height of 1920×1080, shape type code 1, and the actual pixel size of the outline of 192×192, is then stored in the labs array.
[0038] S5: Batch Marking and Data Upload. Repeat steps S2-S4 to complete the marking operation for the other two rectangular parts in the sample image. The marking data for the three targets is stored in the labs array. All data is based on the transform attribute operation information. After marking is completed, the user clicks the upload button. The front end packages and uploads the sample image and the labs array to the server. The server program parses the data and stores it in the database for subsequent AI model training.
[0039] In step S2, the preset size of the outline graphic can be adjusted in real time by screen touch or mouse operation combined with the transform property. If the preset size is too large or too small, the user can directly adjust it to a suitable size by pinching with two fingers or using the mouse wheel, and then perform position and angle calibration. In step S4, the movement information and scaling information are converted by sc to accurately correspond to the actual position and size of the part in the sample image, ensuring that the marking data is without deviation.
[0040] Specific implementation steps for AI recognition accuracy verification method: The configuration of the labeling system based on image AI recognition training, and the specific process of the AI recognition accuracy verification method are as follows: T1: Import of Sample Image to be Verified and Calculation of Scale Factor. The user clicks the image selection button to import the previously marked industrial part sample image or a new sample image to be verified. The image is displayed in the canvas area. The data processing unit repeats step S1 to calculate the scale factor sc=2, which provides the basis for the subsequent size and position adaptation of the outline graphic.
[0041] T2: Sample Image Upload. When the user clicks the recognition button in the menu area, the front end only uploads the sample image to the server, without carrying the labs array labeling data.
[0042] T3: Sample Image Forwarding. After receiving the data, the server program detects the absence of a labs array, converts the sample image format to RGB channels, normalizes it to 640×640 pixels, and forwards it to the image AI recognition program.
[0043] T4: AI Recognition and Result Feedback. The image AI recognition program performs forward inference on the sample image, identifies three rectangular targets, and generates recognition results. The shape type is 1, and the position coordinates are X=405, Y=302; X=800, Y=450; X=1200, Y=600, respectively. The size of each target is approximately 190×190 pixels. The recognition result also includes the approximate position and size information of the target in the sample image. The result is then fed back to the front end through the server program.
[0044] T5: Contour Graphic Generation and Alignment. After receiving the recognition results, the front-end interaction module automatically calls up the rectangular contour graphic. Based on the scaling factor sc=2, it performs pixel ratio conversion on the transform attribute operation information, adjusts the size and position of the contour graphic to adapt the pixel ratio of the contour graphic to the sample image. At the same time, based on the approximate position and size information in the recognition results, it initially locates the display position and size of the contour graphic, and then performs fine adjustment through the transform attribute to align the outer contour of the contour graphic with the shape of the part in the sample image. The user observes that all three contour graphics completely overlap with the part, with a deviation of ≤2 pixels.
[0045] T6: Accuracy Assessment. Users judge the accuracy of AI recognition by observing the overlap between the outline graphic and the target shape in the sample image. If there are issues such as outline offset or shape misjudgment, the user can return to the labeling step to supplement the labeled data and retrain the AI model until the recognition accuracy meets the requirements.
[0046] In step T5, the initial positioning based on the recognition results can significantly reduce the amount of manual adjustment work. Combined with the scaling factor sc to adapt the pixel dimension, it ensures that the outline graphic display accurately matches the sample image. In step T6, users can intuitively judge the recognition effect and realize the rapid verification of the AI training effect without the need for additional third-party tools.
[0047] The system and method of this embodiment were used to mark and validate AI training on 100 industrial part sample images. Each image contained 2-5 shape targets. The results showed that the marking time for a single image was reduced by more than 60% compared to traditional point-by-point marking, mainly due to the rapid adjustment effect brought by contour graphics replacing point-by-point marking and the transform attribute. The marking data deviation was ≤2 pixels, thanks to the accurate capture of transform attribute operation information and the conversion effect of the scale coefficient sc. The AI recognition accuracy reached more than 98% after 3 rounds of iterative training, achieving a simultaneous improvement in marking efficiency, data accuracy and AI training effect. This meets the actual needs of large-scale image AI recognition training, fully demonstrating the technical advantages of the two core innovations of this invention, and effectively solving the pain points of traditional point-by-point manual marking.
[0048] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.
Claims
1. A labeling system for training AI-based image recognition, characterized in that, The system includes a front-end interaction module and a server module. The front-end interaction module has a menu area, a candidate graphic marker area, and a canvas area. The menu area includes selection, marking, upload, and recognition buttons. The candidate graphic marker area stores various candidate basic shape icons, each corresponding to a transparent outline graphic. These outline graphics replace the traditional point-by-point annotation method to mark target features in sample images. The canvas area displays the sample image and outline graphic, and supports operations on the outline graphic via mouse or touchscreen gestures. The `transform` property allows for real-time 2D / 3D transformation of the outline graphic, enabling rotation, scaling, movement, and tilting. The server module has a server program and an image AI recognition program installed. The server program receives the sample image and operation information based on the `transform` property from the front-end interaction module. The image AI recognition program recognizes the sample image and returns the recognition result. The front-end interaction module and the server module are communicatively connected to transmit the sample image and `transform` property operation information.
2. The image AI recognition training labeling system according to claim 1, characterized in that, The front-end interaction module is also equipped with a data processing unit, which is used to obtain the ratio coefficient sc of the actual pixels of the sample image and the pixels displayed on the screen, and to perform pixel ratio conversion processing on the contour graphic related operation information obtained based on the transform attribute.
3. The image AI recognition training labeling system according to claim 2, characterized in that, The front-end interaction module is also equipped with a storage unit. The storage unit stores the marking data through the labs array. The marking data is based on the operation information of the transform attribute, and also includes the movement, scaling, rotation, and tilt information after conversion by the scaling factor sc, as well as the pixel width and height of the sample image, the shape type and pixel size corresponding to the outline graphic.
4. An image AI recognition and tagging method, characterized in that, A labeling system for image AI recognition training according to any one of claims 1-3 includes the following steps: S1, import sample images from folders or albums using the image selection key, display the sample images on the canvas area, and the data processing unit obtains the scaling factor sc. S2, click the target shape icon in the candidate graphic marker bar, and the corresponding outline graphic will be displayed in the upper center of the canvas area according to the preset size; S3 listens for screen touch or mouse operations and uses the transform property to instantly rotate, scale, move, or tilt the outline graphic, making the outer outline of the outline graphic completely overlap with the target shape in the sample image. S4, click the label button, multiply the original operation information in the transform property by the scaling factor sc and convert it into operation information in the pixel dimension of the sample image, and save it to the labs array; S5. Repeat steps S2-S4 to complete the labeling of multiple targets in the sample image. Click the upload button to upload the sample image and the labs array storing transform attribute operation information to the server module for training the image AI recognition program.
5. The image AI recognition and tagging method according to claim 4, characterized in that, In step S4, the movement information of the transform attribute includes the distance between the center of the outline graphic and the center of the canvas area. After multiplying the distance by the scaling factor sc, it is converted into the pixel distance between the outline graphic and the center of the sample image and saved to the labs array.
6. The image AI recognition and tagging method according to claim 4, characterized in that, In step S4, the scaling information of the transform attribute includes the screen display size of the outline graphic. After multiplying the display size by the scaling factor sc, it is converted into the corresponding pixel size of the outline graphic on the sample image and saved to the labs array.
7. The image AI recognition and tagging method according to claim 4, characterized in that, In step S2, the preset size of the outline graphic is a standard size that adapts to the display ratio of the canvas area, and the size of the outline graphic can be adjusted in real time by screen touch or mouse operation combined with the transform property.
8. A method for verifying the accuracy of image AI recognition, characterized in that, The image AI recognition training labeling system described in any one of claims 1-3 is used. Includes the following steps: T1: Import the sample image to be verified using the image selection key. The sample image is displayed on the canvas area, and the data processing unit obtains the scaling factor sc. T2, Click the recognition button to upload the sample image to the server module, but do not upload the corresponding labeling data; T3, after the server program detects that it has only received sample images, it forwards them to the image AI recognition program; T4, the image AI recognition program, after recognizing the sample image, returns the recognition result including the shape type; T5's front-end interaction module generates a corresponding contour graphic based on the recognition results. The transform property is used to instantly rotate, scale, move, or tilt the contour graphic to align it with the target shape in the sample image. Users can judge the recognition accuracy by observing the overlap between the contour graphic and the target shape.
9. The image AI recognition accuracy verification method according to claim 8, characterized in that, In step T5, the front-end interaction module performs pixel ratio conversion on the operation information of the transform property based on the ratio coefficient sc, and adjusts the size and position of the outline graphic to make the pixel ratio of the outline graphic match that of the sample image.
10. The image AI recognition accuracy verification method according to claim 8, characterized in that, The recognition results also include the approximate position and size information of the target in the sample image. The front-end interaction module initially locates the display position and size of the outline graphic based on the information, and then makes fine adjustments through the transform attribute to make the outline graphic accurately aligned with the target shape.