Device for synthesizing data
By introducing equipment for image filtering, automatic labeling, manual adjustment and update model modules into the intelligent labeling system, the problem of poor model performance caused by small data volume is solved, and the labeling efficiency and accuracy are improved.
Patent Information
- Application Number
- CN202411642075.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-15
- Publication Date
- 2025-03-11
AI Technical Summary
The existing intelligent annotation system faces the difficulty in data acquisition caused by complex scenarios and harsh environments, resulting in small data volume and poor model performance during automatic annotation.
Provided is a device for synthesising data, including a picture filtering module, an automatic labeling module, a manual adjustment module and a model update module. By filtering out easy-to-marked and difficult-to-marked pictures, using the automatic labeling module to quickly label easy-to-marked pictures, using the manual adjustment module to correct difficult-to-marked pictures, and expand the training set by updating the model module to retrain the model.
It improves image labeling efficiency, reduces the workload of manual labeling, improves the accuracy and reliability of automatic labeling, and can adapt to more types of image data.
Smart Images
Figure CN119600379A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of synthetic data, and in particular to a device for synthesizing data. Background Art
[0002] Data is the core driving force of artificial intelligence algorithms. Supervised learning algorithms require a large amount of labeled data as data sets for training, verification and testing of neural networks. Data labeling is the process of processing unprocessed initial data with the help of labeling tools and converting it into machine-readable information.
[0003] Since data in different fields have different characteristics, a lot of research has been done on intelligent labeling tasks in different fields at home and abroad to replace the labeling actions that labelers have to repeat and improve labeling efficiency. For example, in the field of intelligent driving, there are intelligent labeling systems such as Baidu Brain's Easy DL and Tesla's automatic labeling system, which can automatically label pedestrians, vehicles, etc. However, the existing intelligent labeling systems, when encountering data collection difficulties caused by complex scenes and harsh environments, resulting in small data volumes, will cause models trained based on small data volumes to perform poorly in automatic labeling.
[0004] Therefore, it is necessary to provide a new device for synthesizing data to solve the above technical problems. Summary of the invention
[0005] In order to solve the above technical problems, the present invention provides a device for synthesizing data.
[0006] The device for synthesizing data provided by the present invention comprises:
[0007] Image screening module: used to divide the data to be labeled into difficult-to-label images and easy-to-label images. The difficult-to-label images enter the manual adjustment module, and then enter the model update module to expand the training set and retrain the network. The easy-to-label images enter the automatic labeling module;
[0008] Automatic labeling module: used to automatically label the easy-to-label images in the image screening module, and the labeling results are in json format files;
[0009] Manual adjustment module: used to receive difficult-to-label images from the image screening module and implement label modification, creation of new labels, creation of label areas, and deletion of label areas;
[0010] Update model module: select image segmentation model for image screening or automatic labeling;
[0011] Image generation module: used to convert the annotation results of the manual adjustment module and the annotation results of the automatic annotation module into pnng format images that can be used for neural network input.
[0012] Furthermore, the image screening module screens out samples that are difficult to label as shown in the following formula:
[0013]
[0014] Where: A* is the difficult-to-label sample selected by the query strategy; argmin is the minimum; A is all the sample data to be labeled; i is the number of identified targets; c1 and c2 are constants, indicating the weights; ε is the error generated in model prediction or sampling, which obeys the normal distribution.
[0015] Furthermore, the automatic annotation method of the automatic annotation module is as follows:
[0016] (1) Use the image segmentation model to predict the image to be labeled. The target area in the prediction result is the labeled area, and the corresponding category is the data label;
[0017] (2) Use the maximum connected region method to extract the boundary of the target area;
[0018] (3) Selecting a subset of pixels at equal intervals from the extracted boundary point set;
[0019] (4) The pixel subsets selected from each region and the corresponding category labels are saved as a json file, i.e., the automatic annotation results.
[0020] Furthermore, in step (1), the image segmentation model selects the BiSeNet semantic segmentation model to label the target area.
[0021] Furthermore, the loss function of the BiSeNet semantic segmentation model is
[0022]
[0023] Where: L is the main loss; α is the weight used to balance the weights of the main loss and the auxiliary loss; Li is the loss of the i-th stage in the context path; K is the number of all stages in the context path.
[0024] Furthermore, step (1) also includes a Yolact instance segmentation network processed by the BiSeNet semantic segmentation model, which is used to automatically label overlapping images.
[0025] Furthermore, the loss function of the Yolact instance segmentation network includes the classification task, the bounding box regression task, and the mask prediction task:
[0026] L=α 1 L class +α 2 SmoothL 1 +α 3L mask
[0027] Where: L is the total loss; L class is the multi-classification cross entropy loss in the classification task; SmoothL 1 It corresponds to the bounding box regression task, which mainly measures the difference between the target box and the true value box; L mask is the binary cross entropy loss used in the mask prediction task; α1, α2, and α3 are the weights of the three losses.
[0028] Furthermore, it also includes an image annotation system interaction interface, which includes a menu bar, a file list, a tag list and a workspace.
[0029] Furthermore, the file list includes the original list before screening by the image screening module and the list of easy-to-label images and difficult-to-label images after screening. The workspace is used to display the images and their annotation information being processed by the current image screening module, automatic annotation module and manual adjustment module.
[0030] Compared with the related art, the device for synthesizing data provided by the present invention has the following advantages:
[0031] Beneficial effects:
[0032] 1. The device for synthesizing data uses the automatic annotation module to quickly annotate images that are easy to process, and outputs the annotation results as JSON format files. By updating the model module, it can use new data from the manual adjustment module and the automatic annotation module to expand the training set and retrain the model. By automatically processing images that are easy to annotate and manually correcting images that are difficult to process, the overall image annotation efficiency is greatly improved. It has a certain degree of versatility and can gradually adapt to more types of image data as new data is added and the model is updated.
[0033] 2. The image screening module can divide the data to be labeled into difficult-to-label images and easy-to-label images. In this way, fewer manually labeled samples can be used to achieve ideal model performance, reducing unnecessary workload.
[0034] 3. The automatic labeling module uses instance segmentation technology, which can accurately label even when there are multiple overlapping objects in the image. It solves the challenges faced by traditional bounding box labeling technology when dealing with overlapping objects and improves the accuracy and reliability of labeling. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 A system schematic diagram of a device for synthesizing data provided by the present invention;
[0036] Figure 2A schematic diagram of the Yolact instance segmentation network provided by the present invention;
[0037] Figure 3 A flow chart of a device for synthesizing data provided by the present invention;
[0038] Figure 4 A flowchart diagram of the automatic annotation module provided by the present invention;
[0039] Figure 5 A schematic diagram of the interactive interface of the intelligent image annotation system provided by the present invention. DETAILED DESCRIPTION
[0040] 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.
[0041] Please refer to Figure 1 , Figure 2 , Figure 3 , Figure 4 as well as Figure 5 , Figure 1 A system schematic diagram of a device for synthesizing data provided by the present invention; Figure 2 A schematic diagram of the Yolact instance segmentation network provided by the present invention; Figure 3 A flow chart of a device for synthesizing data provided by the present invention; Figure 4 A flowchart diagram of the automatic annotation module provided by the present invention; Figure 5 A schematic diagram of the interactive interface of the intelligent image annotation system provided by the present invention.
[0042] In the specific implementation process, Figure 1 , 3 As shown in , 4, and 5, the device for synthesizing data includes an image screening module, an update model module, an automatic annotation module, a manual adjustment module, and an image generation module, wherein the image screening module and the automatic annotation module respectively improve the data annotation efficiency from two aspects: reducing the amount of manually annotated data and improving the accuracy of automatic annotation;
[0043] Image screening module: divides the data to be labeled into difficult-to-label images and easy-to-label images. The difficult-to-label images enter the manual adjustment module and then the model update module to expand the training set and retrain the network, which helps to improve the model performance and thus improve the accuracy of the automatic labeling results. The image screening module is crucial for data sets with small data volumes, because it is difficult to obtain a high-precision model using a small data set. However, the accuracy of the automatic labeling results is highly dependent on the accuracy of the model. Therefore, this module can be used to improve the model accuracy and enter the easy-to-label images into the automatic labeling module;
[0044] Update model module: Receive difficult-to-label images from the image screening module, select image segmentation models to screen or automatically label images, and facilitate users to use different models according to their needs;
[0045] Automatic labeling module: automatically labels the easy-to-label images in the image screening module, and the labeling results are in json format files;
[0046] Manual adjustment module: receives the standard results of the automatic labeling module and implements the modification of labels, creation of new labels, creation of labeling areas, and deletion of labeling areas;
[0047] Image generation module: Convert the annotation results of the manual adjustment module and the automatic annotation module into PNG format images that can be used for neural network input.
[0048] The operation steps of the device used to synthesize data are as follows:
[0049] S1: Input data and collect the image dataset to be annotated. The image dataset to be annotated can enter S2 to update the model, or directly jump to S4 for automatic annotation, or jump to S5 for manual annotation;
[0050] S2: Update model. The update model module updates the image screening or automatic labeling model. If no update is required, skip this step.
[0051] S3: Image screening: The image screening module screens the images to be annotated, and the easy-to-annotate images that are screened out enter S4 for automatic annotation, and the difficult-to-annotate images that are screened out jump to S5 for manual annotation;
[0052] S4: Automatic annotation, the automatic annotation module automatically annotates the results of image screening;
[0053] S5: manual adjustment, the manual adjustment module manually annotates and adjusts the image to be annotated or the result of automatic annotation;
[0054] S6: Generate pictures. The picture generation module will mark the json format file and generate png format pictures in batches.
[0055] Image screening module based on active learning:
[0056] The image screening function is realized by using the idea of active learning. Active learning is to screen out the most valuable candidate sets of difficult-to-label samples based on the query strategy. The core of active learning is the query strategy. In order to screen out difficult-to-label samples more comprehensively, a query strategy suitable for image data is designed, as shown in the following formula:
[0057]
[0058] Where: A* is the difficult-to-label sample selected by the query strategy; argmin is the minimum; A is all the sample data to be labeled; i is the number of identified targets; c1 and c2 are constants, indicating the weights; ε is the error generated in model prediction or sampling, which obeys the normal distribution.
[0059] c 1 +c 2 =10 <c 1 , c 2 <1
[0060] ε~N(0,1)
[0061] m i =max(P i1|θ , P i2|θ …, P iN|θ )
[0062] n i =second_max(P i1|θ , P i2|θ …, P iN|θ )
[0063] m i is the maximum value among the selection probabilities;
[0064] n i ——Select the second largest value in the probability; N is the number of categories in the image; θ is the set of initial model parameters; P iN It is the result of normalizing the scores of each category in the image after model prediction and algorithm processing.
[0065] Automatic annotation module based on image segmentation:
[0066] By taking crop lodging area labeling and corn kernel category labeling as examples, we further study the automatic labeling method based on image segmentation:
[0067] (1) Use the image segmentation model to predict the image to be labeled. The target area in the prediction result is the labeled area, and the corresponding category is the data label;
[0068] (2) Use the maximum connected region method to extract the boundary of the target area;
[0069] (3) Selecting a subset of pixels at equal intervals from the extracted boundary point set;
[0070] (4) The pixel subsets selected from each region and the corresponding category labels are saved as a json file, i.e., the automatic annotation results.
[0071] The target area is annotated by selecting the BiSeNet semantic segmentation model. BiSeNet is a bidirectional segmentation network, one branch of which is a spatial branch that can extract and encode rich spatial information, and the other branch uses a fast downsampling strategy to obtain high-quality semantic information and receptive field. The fusion module combines semantic information with spatial information for image prediction.
[0072] The loss function adopts a combination of main and auxiliary methods. The main loss function supervises the output of the entire dual network, and the auxiliary loss function supervises the output of the context path. The loss function is
[0073]
[0074] Where: L is the main loss; α is the weight used to balance the weights of the main loss and the auxiliary loss; L i is the loss of the i-th stage in the context path; K is the number of all stages in the context path.
[0075] Since semantic segmentation divides the entire image into pixel groups and attempts to classify each pixel in the image, instance segmentation can be seen as a combination of target detection and semantic segmentation. It can classify different types of instances and determine boundaries. For labeling overlapping objects, it is necessary not only to classify these different objects, but also to determine the boundaries of the objects. Therefore, instance segmentation is more suitable for automatic labeling of overlapping images.
[0076] In order to ensure the accuracy of automatic annotation results and reduce the workload of manual adjustment, improvements were made on the basis of the Yolact instance segmentation network. The model used for corn kernel image annotation uses Resnet50 as the basic network to construct a feature pyramid, output multi-scale feature layers p3~p7, and form two branches after the feature pyramid. Branch 1 is used to complete three tasks: multi-scale object detection, bounding box regression, and mask coefficient prediction, and perform non-maximum suppression algorithm processing; Branch 2 generates a template mask based on feature layer p3, then combines the mask coefficient and template mask, and obtains the final mask after bounding box cropping. Finally, the network outputs the full-image mask, target box, and classification results; Yolact's loss function includes three parts: classification task, bounding box regression task, and mask prediction task:
[0077] L=α 1 L class +α 2 SmoothL 1 +α 3 L mask
[0078] Where: L is the total loss; L class is the multi-classification cross entropy loss in the classification task; SmoothL 1 It corresponds to the bounding box regression task, which mainly measures the difference between the target box and the true value box; L mask is the binary cross entropy loss used in the mask prediction task; α1, α2, and a3 are the weights of the three losses.
[0079] Since corn kernels are small in size and have relatively fixed scale changes, in order to more accurately distinguish the types of corn kernels, the feature pyramid structure of Yolact is improved, such as Figure 2 As shown, a suitable model prior frame is fitted for corn kernels.
[0080] Improvement on feature pyramid: The feature pyramid integrates deep and shallow features and has different receptive fields. A large receptive field is conducive to detecting large objects, while a small receptive field is conducive to detecting small objects. Therefore, to achieve high-precision detection of corn kernels, a small receptive field should be given for corn kernels while integrating deep and shallow features. Figure 2 As shown in the figure, the feature pyramid will not downsample after outputting the feature layer p5, but will directly upsample and perform feature fusion with the feature layer corresponding to the feature pyramid. The output feature layers p3 and p4 have both fusion features and small receptive fields, which can greatly improve the detection accuracy of corn kernels.
[0081] Adapt the appropriate model prior frame: Yolact detects targets based on the prior frame. Since the scale of the classified objects in the scene is relatively fixed, the size and aspect ratio of the prior frame are redesigned for the classified objects to make the prior frame consistent with the detected object, which helps to improve the detection accuracy. According to the application scenario, the aspect ratio of the bounding box is set to 1:1, 1:3, and 3:1. The basic side lengths of the three prior frames are 12, 24, and 48 respectively. The basic side lengths are adjusted according to different ratios to ensure that the area of the prior frame is equal. The ratio of 1:1 is mainly used to detect corn kernels and corn pith, and the ratios of 1:3 and 3:1 are used to detect corn stalks.
[0082] Use the maximum connected component method to extract the boundary of the target region:
[0083] 1. Binarization: Convert the predicted mask image into a binary image, that is, set the pixels belonging to a specific category to white (255) and the rest of the pixels to black (0);
[0084] 2. Connected domain analysis: Use the connected domain analysis algorithm to find the largest connected area. This step is used to extract the largest single category area from the binary image.
[0085] 3. Extract boundaries: Perform edge detection on the largest connected area. You can use Canny edge detection or other edge detection algorithms to obtain the boundary contour of the area.
[0086] In the extracted boundary point set, select a subset of pixels at equal intervals:
[0087] 1. Calculate the spacing: Determine a reasonable spacing value based on the length of the boundary contour to ensure that the selected pixel points can better represent the boundary shape;
[0088] 2. Equally spaced sampling: Pixel points are selected at equal intervals along the boundary contour to form a subset of pixels, which will serve as the key points of the boundary contour.
[0089] Results and Analysis:
[0090] 1. Functional verification of intelligent image annotation system
[0091] The effect of the interactive interface of the intelligent image annotation system is demonstrated, and the effectiveness of the image screening strategy and automatic annotation method proposed in the present invention is verified through experiments.
[0092] Intelligent image annotation system:
[0093] The interactive interface of the intelligent image annotation system includes a menu bar, a file list, a tag list, and a workspace. The file list includes the original list before screening by the image screening module and the list of easy-to-annotate images and difficult-to-annotate images after screening. The workspace is used to display the images and their annotation information being processed by the current image screening module, automatic annotation module, and manual adjustment module. It can realize five functions: image screening, automatic annotation, model update, manual adjustment, and image generation.
[0094] Image filtering function verification:
[0095] The proposed query strategy is compared with the minimum confidence query strategy and the edge sampling query strategy, which proves that the proposed method is more effective. 50 pictures are given for the two scenes of crop lodging area labeling and corn kernel category labeling, and the results of using the minimum confidence query strategy, edge sampling query strategy and the proposed query strategy are compared. Among them, the probability screening threshold set by the minimum confidence query strategy is 0.83; the probability threshold set by the edge sampling query strategy is 0.75; the query strategy proposed by the present invention is based on multiple experiments in the model to find difficult-to-label samples, obtain empirical values, and set the parameters to c1=0.55 and c2=0.45, as shown in Table 1:
[0096] Table 1 Comparison of screening quantity
[0097]
[0098] When faced with difficult-to-label samples, there are cases where the results of automatic labeling of images screened using the minimum confidence query strategy and the results of automatic labeling of images screened using the edge sampling query strategy are not screened out. The method proposed in the present invention can screen out these difficult-to-label samples, which can prove that the method proposed in the present invention can screen more comprehensively.
[0099] Automatic annotation function verification:
[0100] The results of automatic labeling by the image labeling system verify that the improved Yolact instance segmentation method can better label overlapping images. The intelligent image labeling system of the present invention can realize the automatic labeling function, and can obtain good labeling results for the automatic labeling of crop lodging pictures. The automatic labeling results based on the improved Yolact instance segmentation method are more accurate than those based on the BiSeNet semantic segmentation method, and can label the overlapping corn pith, corn kernels, etc. in the corn kernel image. Therefore, the present invention proposes that the use of instance segmentation method for labeling is effective for overlapping images.
[0101] The intelligent labeling system proposed in the present invention can be applied to the automatic labeling of various scenes, and can automatically label scenes with small data volumes.
[0102] 2. Comparative experiment
[0103] The results of manual and intelligent annotation of corn kernel images were quantitatively compared to verify the advantages of the intelligent image annotation system proposed in the present invention. At the same time, the system was compared with the existing system. The results show that the intelligent image annotation system proposed in the present invention is not only simple to operate, but also can greatly reduce the annotation time and improve the annotation efficiency.
[0104] Quantitative comparison:
[0105] In order to prove that the proposed intelligent image annotation system improves the annotation efficiency, 10 corn kernel pictures were randomly selected, and the manual correction ratio, time comparison and accuracy comparison experiments were conducted for manual annotation and intelligent annotation, as shown in Table 2:
[0106] Table 2 Proportion of manual correction
[0107]
[0108] The proportion of manual correction reflects the labor cost consumption. The average proportion of manual correction is calculated to be 7.14%, while the proportion of manual correction of manual annotation is 100%. Therefore, the intelligent image annotation proposed in the present invention can significantly reduce the labor cost of annotation.
[0109] The manual annotation time of 5 inexperienced annotators and the time taken by the intelligent image annotation system were counted. The intelligent image annotation time includes the time for manual correction after automatic annotation, as shown in Table 3:
[0110] Table 3 Comparison of time between manual labeling and intelligent labeling
[0111]
[0112] The comparison results are shown in Table 3, and the time reduction rate of intelligent labeling is more than 70%.
[0113] The accuracy of the manual and intelligent annotation results is compared with the true value of the professional annotation, and the miou value is used as the standard to measure the accuracy. The comparison results are shown in Table 4:
[0114] Table 4 Comparison of accuracy between manual labeling and intelligent labeling:
[0115] Annotation method Average miou value / % Manual annotation 88 Smart Labeling 91
[0116] It can be seen that the accuracy of intelligent labeling results is higher than that of manual labeling.
[0117] Compared with existing intelligent labeling systems:
[0118] Compared with Baidu Brain's Easy DL intelligent labeling system, taking the lodging scene as an example, the comparison is made from the following aspects:
[0119] (1) Is the automatic labeling process simple and convenient?
[0120] The steps for automatic annotation using the Easy DL intelligent annotation system are as follows:
[0121] S1: Select the task type as needed: active learning, specified model;
[0122] S2: When "Active Learning" is selected, select the data set to be labeled, and four rounds of difficult examples will be screened, and then jump to S3; when "Specify Model" is selected, the system's specified model will be used for automatic labeling:
[0123] S3: After screening out samples that are difficult to label, they need to be manually labeled online on the platform;
[0124] S4: Send the manually annotated hard example labels to the learner training;
[0125] S5: The learner automatically labels the remaining samples based on the learned model.
[0126] Compared with the automatic annotation process of the intelligent image annotation system proposed in this invention, the active learning process of Easy DL requires multiple rounds of batch screening, manual confirmation and manual annotation, which is cumbersome, time-consuming and difficult to operate.
[0127] (2) Does the system have the function of changing models?
[0128] When selecting the "Specify Model" function of Easy DL, you can select the system's internal model (the model that has been trained by Easy DL and BML) or the model obtained through active learning on the system, but you cannot import external models; the "Update Model" function of the intelligent image annotation system proposed in the present invention can import any image segmentation model, which is convenient for users to import specified models according to their needs.
[0129] (3) Comparison of labeling time
[0130] EasyDL and the system of the present invention were used to annotate 50 pictures of lodging scene data respectively, and the annotation time was compared. In the Easy DL system, the "active learning" function was first used to screen difficult-to-annotate samples and perform online manual annotation, and then automatic annotation was performed, which took a total of 35 minutes. Using the intelligent image annotation system of the present invention, the difficult-to-annotate samples were first screened and manually annotated, and then automatic annotation was performed, which took a total of 17 minutes.
[0131] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a general hardware platform, and of course, by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the relevant technology can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.
[0132] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic features of the present invention. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the attached claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention, and any figure mark in the claims should not be regarded as limiting the claims involved.
[0133] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.
Claims
1. A device for synthesizing data, characterized in that include: Image screening module: used to divide the data to be labeled into difficult-to-label images and easy-to-label images. The difficult-to-label images enter the manual adjustment module, and then enter the model update module to expand the training set and retrain the network. The easy-to-label images enter the automatic labeling module; Automatic labeling module: used to automatically label the easy-to-label images in the image screening module, and the labeling results are in json format files; Manual adjustment module: used to receive difficult-to-label images from the image screening module and implement label modification, creation of new labels, creation of label areas, and deletion of label areas; Update model module: select image segmentation model for image screening or automatic labeling; Image generation module: used to convert the annotation results of the manual adjustment module and the annotation results of the automatic annotation module into PNG format images that can be used for neural network input.
2. The device for synthesizing data according to claim 1, characterized in that The image screening module screens out samples that are difficult to label as shown in the following formula: Where: A* is the difficult-to-label sample selected by the query strategy; argmmin is the minimum; A is all the sample data to be labeled; i is the number of identified targets; c1 and c2 are constants, indicating the weights; ε is the error generated in model prediction or sampling, which obeys the normal distribution.
3. The device for synthesizing data according to claim 2, characterized in that The automatic annotation method of the automatic annotation module is as follows: (1) Use the image segmentation model to predict the image to be labeled. The target area in the prediction result is the labeled area, and the corresponding category is the data label; (2) Use the maximum connected region method to extract the boundary of the target area; (3) Selecting a subset of pixels at equal intervals from the extracted boundary point set; (4) The pixel subsets selected from each region and the corresponding category labels are saved as a json file, i.e., the automatic annotation results.
4. The device for synthesizing data according to claim 3, characterized in that In step (1), the image segmentation model selects the BiSeNet semantic segmentation model to label the target area.
5. The device for synthesizing data according to claim 4, characterized in that The loss function of the BiSeNet semantic segmentation model is Where: L is the main loss; α is a weight used to balance the weights of the main loss and the auxiliary loss; L i is the loss of the i-th stage in the context path; K is the number of all stages in the context path.
6. The device for synthesizing data according to claim 5, characterized in that Step (1) also includes a Yolact instance segmentation network processed by the BiSeNet semantic segmentation model, which is used to automatically annotate overlapping images.
7. The device for synthesizing data according to claim 6, characterized in that The loss function of the Yolact instance segmentation network includes classification tasks, bounding box regression tasks, and mask prediction tasks: L=α1L class +α2SmoothL1+α3L mask Where: L is the total loss; L class It is the multi-classification cross entropy loss in the classification task; SmoothL1 corresponds to the bounding box regression task, which mainly measures the difference between the target box and the true value box; L mask is the binary cross entropy loss used in the mask prediction task; α1, α2, and α3 are the weights of the three losses.
8. The device for synthesizing data according to claim 7, characterized in that It also includes an image annotation system interactive interface, which includes a menu bar, a file list, a tag list, and a workspace.
9. The device for synthesizing data according to claim 8, characterized in that The file list includes the original list before screening by the image screening module and the list of easy-to-label images and difficult-to-label images after screening. The workspace is used to display the images and their annotation information being processed by the current image screening module, automatic annotation module, and manual adjustment module.