Data cleaning method and device, electronic equipment and storage medium

Abstract

Embodiments of the present application disclose a data cleaning method and device, electronic equipment and storage medium. The method comprises: when a data cleaning request for a target model identifier is received, obtaining a data set corresponding to the target model identifier, the data set comprising annotation data and model prediction data of multiple images; for each of the images, comparing the annotation data and the model prediction data, and adding an error mark to the annotation data according to a comparison result; displaying the images with the error mark added, obtaining a returned image from the displayed images, and returning the returned image to a data quality inspection node, so that the data quality inspection node modifies the annotation data in the returned image according to the error mark. The embodiments of the present application reduce the operation of the user, can improve the data cleaning efficiency, and can display the images with the error mark added, so as to facilitate the user to check whether the mark is correct, and can improve the accuracy of data cleaning.

Description

TECHNICAL FIELD

[0001] Embodiments of the present application relate to the technical field of data processing, in particular to a data cleaning method and device, electronic equipment and storage medium. BACKGROUND

[0002] Before model training, in order to ensure the accuracy of the labeled data, the labeled data needs to be cleaned.

[0003] In the prior art, the data cleaning program is executed by an algorithm personnel, the data cleaning program generates a fixed format json file including the cleaning result, and the algorithm personnel uploads the json file to a platform, and the platform program parses the file content, supplements all the image annotation data in the file with error marks and returns them to the quality inspector for re-audit.

[0004] The above operation mode of generating an error file by an algorithm personnel and then uploading it to the platform for return needs to consume multiple human resources, and cannot check whether the error marks for the annotation data in the file are really wrong, which affects the cleaning accuracy and efficiency. SUMMARY

[0005] Embodiments of the present application provide a data cleaning method and device, electronic equipment and storage medium, which help to improve the data cleaning efficiency and accuracy.

[0006] To solve the above problems, in a first aspect, embodiments of the present application provide a data cleaning method, comprising:

[0007] When a data cleaning request for a target model identifier is received, a data set corresponding to the target model identifier is obtained, and the data set includes annotation data and model prediction data of multiple images;

[0008] For each of the images, the annotation data and the model prediction data are compared, and error marks are added to the annotation data according to the comparison result;

[0009] The images with the error marks are displayed, a returned image is obtained from the displayed images, and the returned image is returned to a data quality inspection node, so that the data quality inspection node modifies the annotation data in the returned image according to the error marks.

[0010] In a second aspect, embodiments of the present application provide a data cleaning device, comprising:

[0011] A data set obtaining module is configured to, when a data cleaning request for a target model identifier is received, obtain a data set corresponding to the target model identifier, and the data set includes annotation data and model prediction data of multiple images;

[0012] a data comparison module configured to compare the labeled data and the model prediction data for each of the images, and add an error mark in the labeled data according to a comparison result;

[0013] a data return module configured to display the images with the error marks, obtain a returned image from the displayed images, and return the returned image to the data quality inspection node, so that the data quality inspection node modifies the labeled data in the returned image according to the error mark.

[0014] In a third aspect, an electronic device is provided, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the data cleaning method according to the embodiments of the present application when executing the computer program.

[0015] In a fourth aspect, a computer readable storage medium is provided, which stores a computer program, and the program is executable on a processor to implement the data cleaning method according to the embodiments of the present application.

[0016] The data cleaning method, device, electronic device, and storage medium provided by the embodiments of the present application can achieve direct comparison of the labeled data and the model prediction data, reduce user operations, improve data cleaning efficiency, and display images with error marks, so as to facilitate user checking of whether the marks are correct, thereby improving the accuracy of data cleaning. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0018] Figure 1 is a flowchart of a data cleaning method provided by the embodiments of the present application;

[0019] Figure 2 is a flowchart of another data cleaning method provided by the embodiments of the present application;

[0020] Figure 3 is a flowchart of data comparison in the embodiments of the present application;

[0021] Figure 4 is a flowchart of data back in the embodiments of the present application;

[0022] Figure 5 is a structural schematic diagram of a data cleaning device provided by the embodiments of the present application;

[0023] Figure 6 is a structural schematic diagram of an electronic device provided by the embodiments of the present application. DETAILED DESCRIPTION

[0024] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0025] Figure 1 is a flowchart of a data cleaning method provided by the embodiments of the present application. The data cleaning method can be executed by a computer or other electronic device, as shown in Figure 1 , the method comprises the following steps:

[0026] Step 110, when a data cleaning request for a target model identifier is received, a data set corresponding to the target model identifier is obtained, and the data set comprises annotation data and model prediction data of multiple images.

[0027] Multiple task identifiers to be cleaned can be displayed. When a user clicks a target task identifier, model identifiers under the target task identifier can be displayed. When the user clicks a target model identifier under the target task identifier, it is determined that a data cleaning request for the target model identifier is received. At this time, all data sets corresponding to the target model identifier are obtained, and each data set comprises multiple images, annotation data and model prediction data of each image.

[0028] Step 120, for each of the images, the annotation data and the model prediction data are compared, and according to the comparison result, an error mark is added in the annotation data.

[0029] For each data set, the labeled data and the model prediction data corresponding to each image in the data set are found, and the labeled data and the model prediction data of each image in each data set are compared respectively to determine whether the labeled data and the model prediction data of an image are consistent. If the labeled data and the model prediction data are inconsistent, an error mark is added to the labeled data of the image.

[0030] In step 130, the image with the added error mark is displayed, a returned image is obtained from the displayed image, and the returned image is returned to the data quality inspection node, so that the data quality inspection node modifies the labeled data in the returned image according to the error mark.

[0031] After the comparison of the target model identification corresponding data set is completed, the image with the added error mark is displayed to facilitate the user to check whether the added error mark is correct. If the user finds that there is an obvious comparison error problem, for example, the labeled data of an image actually has no error but an error mark is added through comparison, the comparison algorithm accuracy is improved by returning to the algorithm personnel, and then the target model identification corresponding data set is cleaned again. If the user believes that the image with the added error mark has no obvious comparison error problem, the returned operation can be performed. In the displayed image, the user can select the image to be returned, so that the electronic device performing the data cleaning method determines the image selected by the user as a returned image, and returns the returned image to the data quality inspection node based on the returned instruction of the user, so that the data quality inspection personnel modify the labeled data in the returned image according to the error mark in the returned image, or the data quality inspection personnel can also compare the labeled data and the model prediction data of the returned image. If the model prediction data has a problem, the target model can be retrained. After the data quality inspection personnel modify the labeled data, the returned image and the labeled data are submitted again, so that the above steps 110 to 130 can be executed again until there is no error labeled data, and the cleaning is completed.

[0032] The data cleaning method provided in the embodiment, by receiving the data cleaning request of the target model identification, obtaining the data set corresponding to the target model identification, comparing the labeled data and the model prediction data for each image, adding an error mark to the labeled data according to the comparison result, displaying the image with the added error mark, obtaining a returned image from the displayed image, and returning the returned image to the data quality inspection node, so that the data quality inspection node modifies the labeled data in the returned image according to the error mark, the comparison of the labeled data and the model prediction data is realized, the operation of the user is reduced, the data cleaning efficiency can be improved, and the image with the added error mark can be displayed to facilitate the user to check whether the mark is correct, so that the accuracy of data cleaning can be improved.

[0033] In an embodiment of the present application, the comparison of the annotation data and the model prediction data for each of the images includes:

[0034] For each of the images, the annotation detection frame in the annotation data is traversed, and a prediction detection frame of the same type as the annotation detection frame is searched in the model prediction data. If the annotation detection frame and the prediction detection frame are inconsistent, the error mark is added in the annotation data corresponding to the annotation detection frame.

[0035] The prediction detection frame in the model prediction data is traversed. If the prediction detection frame does not exist in the annotation data, the error mark corresponding to the prediction detection frame is added in the annotation data.

[0036] The target model can be a target detection model, the annotation data includes an annotation detection frame and a corresponding detection frame type, and the model prediction data includes a prediction detection frame and a corresponding detection frame type.

[0037] For each image, the annotation detection frame in the annotation data can be traversed, the prediction detection frame of the same type as the annotation detection frame can be searched in the model prediction data, and the annotation detection frame and the prediction detection frame can be compared to determine whether the annotation detection frame and the prediction detection frame are consistent. If the annotation detection frame and the prediction detection frame are consistent, it is determined that the annotation detection frame is correct. If the annotation detection frame and the prediction detection frame are inconsistent, it is determined that the annotation detection frame has a multi-label error, an error mark of the error type of multi-label is generated, and the error mark is added in the annotation data corresponding to the annotation detection frame. After the annotation detection frame is traversed, the prediction detection frame is traversed, and the annotation detection frame of the same type as the prediction detection frame is searched in the annotation data. If the annotation detection frame of the same type as the prediction detection frame cannot be found in the annotation data, it is determined that the annotation data does not have the annotation detection frame of the same type as the prediction detection frame, an error mark of the error type of missing label is generated, and the error mark is added in the annotation data at the position corresponding to the prediction detection frame.

[0038] The annotation detection frame and the prediction detection frame are compared by traversal, which can avoid omission and improve the accuracy of comparison of the annotation detection frame and the prediction detection frame.

[0039] In an embodiment of the present application, if the labeled detection frame is inconsistent with the predicted detection frame, the error mark is added to the labeled detection frame in the labeled data, including: determining the intersection over union of the labeled detection frame and the predicted detection frame; if the intersection over union is greater than the intersection over union threshold corresponding to the target model identifier, it is determined that the labeled detection frame is inconsistent with the predicted detection frame, and the error mark is added to the labeled detection frame in the labeled data.

[0040] After traversing the labeled detection frame, finding the predicted detection frame of the same type as the labeled detection frame from the model prediction data, determining the intersection and union of the labeled detection frame and the predicted detection frame, and determining the ratio of the intersection to the union, the intersection over union (IoU) of the labeled detection frame and the predicted detection frame is obtained; the intersection over union is compared with the intersection over union threshold corresponding to the target model identifier, if the intersection over union is less than or equal to the intersection over union threshold, it is determined that the labeled detection frame is consistent with the predicted detection frame, and the labeled detection frame is determined to be a correct detection frame, if the intersection over union is greater than the intersection over union threshold, it is determined that the labeled detection frame is inconsistent with the predicted detection frame, then the error mark of multiple labels is generated, and the error mark is added to the labeled detection frame in the labeled data. By determining whether the labeled detection frame and the predicted detection frame are consistent through the intersection over union, the correctness of the comparison can be ensured under the condition of allowing a certain deviation.

[0041] In an embodiment of the present application, the rejected image is obtained from the displayed image, including:

[0042] In response to a full rejection instruction for a target image in the displayed image, a target batch to which the target image belongs is obtained, the state of the target image in the target batch is set to an unqualified state, the state of the target batch is set to a to-be-modified state, and all images in the target batch are taken as the rejected image; or

[0043] In response to a partial rejection instruction for a target image in the displayed image, a target batch to which the target image belongs is obtained, the state of the target image in the target batch is set to an unqualified state, the state of the target batch is set to a to-be-modified state, and the target image is taken as the rejected image.

[0044] For the displayed image with error label, the user can choose to return all or part. When receiving the user's all return instruction of the target image in the displayed image, the target batch to which the target image belongs in the data set is acquired, the state of the target image in the target batch is set to unqualified state, the state of the target batch is set to to-be-modified state, and all images in the target batch are returned to the data quality inspection node as return images. The data quality inspection node displays the state of the target batch as to-be-modified state based on the instruction of the data quality inspection personnel. After the data quality inspection personnel clicks the target batch, all images in the target batch and the state corresponding to each image are displayed. The user can modify the annotation data of the image in the unqualified state, and can also view the annotation data of other images in the batch.

[0045] When receiving the user's part return instruction of the target image in the displayed image, the target batch to which the target image belongs in the data set is acquired, the state of the target image in the target batch is set to unqualified state, and the state of the target batch is set to to-be-modified state. The target image in the target batch is returned to the data quality inspection node as a return image, and other qualified images in the target batch do not need to be returned to the data quality inspection node. The data quality inspection node displays the state of the target batch as to-be-modified state based on the instruction of the data quality inspection personnel. After the data quality inspection personnel clicks the target batch, the target image in the target batch and the unqualified state corresponding to the target image are displayed. The user can modify the annotation data of the target image in the unqualified state. When there are multiple target images with error labels in one target batch, multiple target images with error labels can be returned as return images at the same time and returned to the data quality inspection node.

[0046] Through the displayed image, the user can check whether the error label of each image is accurate, and can choose to return all or part, which enriches the data return mode.

[0047] In an embodiment of the present application, the target model identifier corresponding data set is acquired, and if the model prediction data is missing in the data set, a prompt information of missing model prediction data is displayed. In response to the model prediction instruction of the data set, the target model corresponding to the target model identifier is used to process multiple images in the data set respectively to obtain model prediction data of each image.

[0048] When the data cleaning request for the target model identifier is received, the data set corresponding to the target model identifier can be obtained from the database, and it is detected whether the model prediction data is missing in the obtained data set. If the model prediction data is missing in the data set, a prompt information of missing model prediction data is displayed to prompt the user which data set is missing the model prediction data. Based on the model prediction instruction of the user to the data set, the model prediction page can be jumped to, and based on the target model and the data set selected by the user, the target model is used to process each image in the data set to obtain the model prediction data of each image in the data set.

[0049] By detecting whether the model prediction data is missing in the data set when the data set corresponding to the target model identifier is obtained, the problem that the model prediction data cannot be labeled and compared with the model prediction data due to the missing model prediction data can be avoided, so that the efficiency of data cleaning can be further improved.

[0050] Figure 2 is a flowchart of another data cleaning method provided by the embodiment of the present application. The data cleaning method can be executed by a computer and other electronic devices, as shown in Figure 2 The method comprises the following steps:

[0051] In step 210, when the data cleaning request for the target model identifier is received, the data set corresponding to the target model identifier is obtained, and the data set comprises the labeled data and the model prediction data of a plurality of images.

[0052] In step 220, for each image, the labeled data and the model prediction data are compared, and an error mark is added in the labeled data according to the comparison result.

[0053] In step 230, the number of error marks in each image is counted, and an error list comprising the images with error marks and the number is generated.

[0054] For each image in each data set, the number of error marks is counted respectively, the total number of error marks can be counted, the number of error marks of the error type of multi-labeling can be counted, and the number of error marks of the error type of missing labeling can be counted. The images with error marks and the number of error marks in each image are generated to generate an error list.

[0055] In step 240, the error list is displayed, the returned image is obtained from the displayed image, and the returned image is returned to the data quality inspection node, so that the data quality inspection node modifies the labeled data in the returned image according to the error mark.

[0056] The error list is displayed, and the image with the error mark and the number of error marks corresponding to the image are displayed in the error list, so as to facilitate the user to check the image with the error mark, and the user can select the image to be returned from the displayed image. After obtaining the image to be returned selected by the user, the image to be returned is returned to the data quality inspection node, so that the data quality inspection node modifies the annotation data in the image to be returned according to the error mark.

[0057] In an embodiment of the present application, after displaying the error list, the method further comprises: in response to a viewing instruction for a target image in the error list, displaying the target image and displaying the annotation data and the error mark corresponding to the annotation data in the target image.

[0058] After displaying the error list, the user can click to view the image with the error mark. When a viewing instruction for a target image selected by the user in the error list is received, the target image is displayed, and the annotation data and the error mark corresponding to the annotation data are displayed in the displayed target image, that is, the annotation detection box and the error mark corresponding to the annotation detection box can be displayed in the displayed target image. By displaying the annotation data and the error mark corresponding to the annotation data in the target image based on the viewing instruction of the user for the target image, the user can view whether the error mark is correct, so as to further improve the data cleaning efficiency.

[0059] The data cleaning method provided in the embodiment can facilitate the user to view the image with the error mark and the number of corresponding error marks by adding the error mark in the annotation data, generating the error list including the image with the error mark and the number of error marks, and displaying the error list, so that the image to be returned can be quickly selected, and the efficiency of data cleaning can be further improved.

[0060] Figure 3 is a flowchart of data comparison in the embodiment of the present application, as Figure 3 shown, the comparison between the annotation data and the model prediction data can be performed by the following steps:

[0061] Step 31, receiving a data cleaning request for a target model identifier under a target task identifier;

[0062] Step 32, obtaining all data sets corresponding to the target model identifier;

[0063] Step 33, judging whether all the data sets have model prediction data of the target model corresponding to the target model identifier, if not, executing step 34, if yes, executing step 36;

[0064] Step 34, reporting an error prompt for the data set lacking the model prediction data;

[0065] Step 35, jump to the model prediction page, perform data prediction based on the selected data set and target model, generate model prediction data, and then perform step 31;

[0066] Step 36, obtain the intersection over union threshold value of the target model;

[0067] Step 37, obtain the labeled data of each data set and the model prediction data;

[0068] Step 38, generate error markers according to the comparison rules;

[0069] Step 39, add the error markers to the labeled data of the image and create an error image list.

[0070] The error image list is the error list described above, which includes images with error markers and the number of corresponding error markers.

[0071] The specific content of each step can be referred to the description of the above embodiments, which will not be repeated here.

[0072] Figure 4 is a flowchart of performing data return in the embodiment of the present application, as shown in Figure 4 The operation of data return can be performed by the following steps:

[0073] Step 41, display the error image list;

[0074] Step 42, determine whether the error markers in the image are correct, if not, perform step 43, if correct, perform step 45;

[0075] Step 43, generate error feedback information of the error markers, and feed back the error feedback information to the algorithm personnel node, so as to improve the accuracy of the data comparison algorithm by the algorithm personnel;

[0076] Step 44, re-perform data cleaning;

[0077] Step 45, obtain the returned image, and return the returned image to the data quality inspection node;

[0078] Step 46, obtain the modified labeled data at the data quality inspection node and submit;

[0079] Step 47, re-perform data cleaning until there is no error in the labeled data.

[0080] The flow of Figure 3 and Figure 4 is executed again, based on Figure 3 and Figure 4The data cleaning operation is performed in the flow of the data set, until there is no error labeled data, that is, until there is no labeled data inconsistent with the model predicted data, and the data cleaning ends.

[0081] The operation of data comparison is integrated into the platform, reducing the manual operation of the algorithm personnel,

[0082] The embodiment of the present application can more simply, quickly and accurately complete the cleaning of the labeled data. In the prior art, the algorithm personnel generates an error file by executing a program and uploads it to the platform to return the operation mode, which needs to consume manpower in many aspects; and it cannot be checked whether the error mark of the labeled data in the file is really wrong, which affects the accuracy of cleaning; and it can only be returned in whole, which affects the efficiency. The embodiment of the present application can directly execute the comparison program of the data on the platform, find out the image data of the comparison error, show the error image list, allow the user to check whether the error mark of each error image is accurate, and can select part of the data to return. Through the data cleaning method of the embodiment of the present application, the cleaning efficiency can be improved by 50%, and the cleaning quality can be improved by 10%.

[0083] Figure 5 is a structural schematic diagram of a data cleaning device provided by the embodiment of the present application, as Figure 5 shown, the device comprises:

[0084] The data set acquisition module 510 is configured to acquire a data set corresponding to a target model identifier when a data cleaning request for the target model identifier is received, and the data set comprises labeled data and model predicted data of a plurality of images.

[0085] The data comparison module 520 is configured to compare the labeled data and the model predicted data for each of the images, and add an error mark in the labeled data according to a comparison result.

[0086] The data return module 530 is configured to show the images with the error mark added, acquire a returned image from the shown images, and return the returned image to a data quality inspection node, so that the data quality inspection node modifies the labeled data in the returned image according to the error mark.

[0087] Optionally, the data comparison module comprises:

[0088] The labeled data traversal unit is configured to traverse a labeled detection frame in the labeled data for each of the images, find a predicted detection frame of the same type as the labeled detection frame in the model predicted data, and add the error mark in the labeled data corresponding to the labeled detection frame if the labeled detection frame is inconsistent with the predicted detection frame.

[0089] The prediction data traversal unit is configured to traverse a prediction bounding box in the model prediction data, and if there is no labeled bounding box of the same type as the prediction bounding box in the labeled data, add an error label corresponding to the prediction bounding box in the labeled data.

[0090] Optionally, the labeled data traversal unit comprises:

[0091] The intersection-over-union ratio determination subunit is configured to determine an intersection-over-union ratio of the labeled bounding box and the prediction bounding box.

[0092] The error label adding subunit is configured to determine that the labeled bounding box and the prediction bounding box are inconsistent if the intersection-over-union ratio is greater than a threshold value of the intersection-over-union ratio corresponding to the target model, and add the error label corresponding to the labeled bounding box in the labeled data.

[0093] Optionally, the apparatus further comprises:

[0094] The error list generation module is configured to count a number of the error labels in each of the images, and generate an error list comprising the images including the error labels and the number.

[0095] The data return module comprises:

[0096] The error list display unit is configured to display the error list.

[0097] Optionally, the data return module further comprises:

[0098] The image information display unit is configured to display the target image in response to a viewing instruction of a target image in the error list, and display the labeled data and the error label corresponding to the labeled data in the target image.

[0099] Optionally, the data return module comprises:

[0100] The all return unit is configured to acquire a target batch to which the target image belongs in response to an all return instruction of a target image in the displayed images, set a state of the target image to an unqualified state in the target batch, set a state of the target batch to a to-be-modified state, and take all images in the target batch as the returned images; or

[0101] The partial return unit is configured to acquire a target batch to which the target image belongs in response to a partial return instruction of a target image in the displayed images, set a state of the target image to an unqualified state in the target batch, set a state of the target batch to a to-be-modified state, and take the target image as the returned image.

[0102] Optionally, the data set acquisition module comprises:

[0103] The promotion information display unit is configured to acquire the data set corresponding to the target model identifier, and display prompt information indicating that the model prediction data is missing if the data set is missing the model prediction data.

[0104] The prediction data acquisition unit is configured to, in response to a model prediction instruction for the data set, process a plurality of images in the data set by using the target model corresponding to the target model identifier, to obtain model prediction data of each of the images.

[0105] The data cleaning device provided by the embodiments of the present application is used to implement the steps of the data cleaning method described in the embodiments of the present application, and the specific implementation manners of the modules of the device are described in the corresponding steps, which will not be repeated here.

[0106] The data cleaning device provided by the embodiments of the present application is used to implement the steps of the data cleaning method described in the embodiments of the present application, and the specific implementation manners of the modules of the device are described in the corresponding steps, which will not be repeated here.

[0107] Figure 6 is a structural schematic diagram of an electronic device provided by the embodiments of the present application, as shown in Figure 6 The electronic device 600 can include one or more processors 610 and one or more memories 620 connected to the processors 610. The electronic device 600 can further include an input interface 630 and an output interface 640 for communicating with another device or system. The program code executed by the processor 610 can be stored in the memory 620.

[0108] The processor 610 in the electronic device 600 calls the program code stored in the memory 620 to execute the data cleaning method in the above-described embodiments.

[0109] According to an embodiment of the present application, a computer readable storage medium is also provided, which includes but is not limited to disk memory, CD-ROM, optical memory, etc. The computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the data cleaning method described in the foregoing embodiments.

[0110] According to one embodiment of the present application, a computer program product is also provided, comprising computer programs or computer instructions, which, when executed by a processor, implement the data cleaning method described in the above embodiments.

[0111] Each of the embodiments in the specification is described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the part of the method embodiments.

[0112] The above describes in detail the data cleaning method, device, electronic device and storage medium provided by the embodiments of the present application. The principle and implementation mode of the present application are described by applying specific examples. The above embodiment description is only used to help understand the method and core idea of the present application. Meanwhile, for the general technical personnel in the art, according to the idea of the present application, the specific implementation mode and application range can be changed. In summary, the content of the specification should not be understood as a limitation of the present application.

[0113] From the above description of the embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software and necessary universal hardware platforms. Of course, it can also be realized by hardware. Based on such understanding, the above technical solutions, essentially or in other words, the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.

Claims

1. A data cleaning method, characterized in that, include: Upon receiving a data cleaning request for a target model identifier, the dataset corresponding to the target model identifier is obtained, and the dataset includes annotation data and model prediction data for multiple images; For each image, the labeled data and the model prediction data are compared, and error markers are added to the labeled data based on the comparison results. The system displays an image with the error markers added, retrieves a returned image from the displayed image, and sends the returned image back to the data quality inspection node so that the data quality inspection node can modify the labeled data in the returned image based on the error markers. The step of obtaining the returned image from the displayed image includes: In response to the rejection instruction for all target images in the displayed images, the target batch to which the target image belongs is obtained, and the status of the target image in the target batch is set to unqualified, the status of the target batch is set to pending modification, and all images in the target batch are used as the rejection images; or In response to a partial rejection instruction for a target image in the displayed images, the target batch to which the target image belongs is obtained, and the status of the target image in the target batch is set to unqualified, the status of the target batch is set to pending modification, and the target image is used as the rejected image. The data quality inspection node modifies the annotation data in the returned image based on the error markers, then resubmits the returned image and the modified annotation data to perform data cleaning until there is no erroneous annotation data.

2. The method according to claim 1, characterized in that, For each image, the labeled data and the model prediction data are compared, and error markers are added to the labeled data based on the comparison results, including: For each image, the annotation detection boxes in the annotation data are traversed, and the prediction detection boxes of the same type as the annotation detection boxes are found in the model prediction data. If the annotation detection boxes and the prediction detection boxes are inconsistent, the error mark is added to the annotation data corresponding to the annotation detection boxes. Traverse the predicted detection boxes in the model prediction data. If there is no labeled detection box of the same type as the predicted detection box in the labeled data, add an error marker corresponding to the predicted detection box in the labeled data.

3. The method according to claim 2, characterized in that, If the labeled detection box is inconsistent with the predicted detection box, then an error marker is added to the labeled data corresponding to the labeled detection box, including: Determine the intersection-union ratio (IUU) of the labeled detection box and the predicted detection box; If the intersection-union ratio is greater than the intersection-union ratio threshold corresponding to the target model identifier, then it is determined that the labeled detection box is inconsistent with the predicted detection box, and the error mark is added to the labeled detection box in the labeled data.

4. The method according to any one of claims 1-3, characterized in that, Before displaying the image with the error marker added, the following is also included: The number of error markers in each image is counted, and an error list is generated including the images with the error markers and the counts. The image displaying the added error marker includes: Display the aforementioned list of errors.

5. The method according to claim 4, characterized in that, Following the display of the error list, the following is also included: In response to a viewing instruction for a target image in the error list, the target image is displayed, and the annotation data and the error markers corresponding to the annotation data are displayed in the target image.

6. The method according to any one of claims 1-3, characterized in that, The step of obtaining the dataset corresponding to the target model identifier includes: Obtain the dataset corresponding to the target model identifier. If the dataset lacks model prediction data, display a prompt message indicating that model prediction data is missing. In response to a model prediction instruction for the dataset, multiple images in the dataset are processed using the target model corresponding to the target model identifier to obtain model prediction data for each image.

7. A data cleaning apparatus, characterized in that, include: The dataset acquisition module is used to acquire the dataset corresponding to the target model identifier when a data cleaning request for the target model identifier is received. The dataset includes annotation data and model prediction data of multiple images. The data comparison module is used to compare the labeled data and the model prediction data for each image, and add error markers to the labeled data according to the comparison results; The data return module is used to display images with the added error markers, obtain return images from the displayed images, and return the return images to the data quality inspection node so that the data quality inspection node can modify the annotation data in the return images according to the error markers. After the data quality inspection node modifies the annotation data in the return images according to the error markers, it resubmits the return images and the modified annotation data to perform data cleaning until there is no erroneous annotation data. The data feedback module includes: The all-return unit is used to respond to the all-return command for the target images in the displayed images, obtain the target batch to which the target image belongs, set the status of the target image in the target batch to an unqualified state, set the status of the target batch to a state to be modified, and use all images in the target batch as the returned images. or The partial return unit is configured to, in response to a partial return instruction for a target image in the displayed images, obtain the target batch to which the target image belongs, set the status of the target image in the target batch to an unqualified state, set the status of the target batch to a state to be modified, and use the target image as the returned image.

8. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the data cleaning method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the data cleaning method according to any one of claims 1 to 6.

Images