Image processing method and device, management system, electronic equipment and storage medium

Abstract

The application provides an image processing method and device, a management system, electronic equipment and a storage medium. The method comprises the following steps: obtaining a to-be-labeled image of a to-be-labeled region; obtaining labeling information of the to-be-labeled image based on a pre-trained labeling model, wherein the labeling model is obtained by pre-training a sample set with the labeling information; determining a target to-be-labeled image whose labeling information does not meet a preset requirement; and correcting the labeling information of the target to-be-labeled image based on the obtained correction information to meet the preset requirement. In this way, the labeling model can be used to label the to-be-labeled image, and the labeling information can be corrected based on this, which can reduce the labeling workload of the workers, improve the labeling efficiency, and ensure the recall rate and accuracy of the final labeling information.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent agriculture, in particular to an image processing method and device, a management system, an electronic device and a storage medium. BACKGROUND

[0002] In the field of agricultural intelligence, in order to grasp the crop growth, field environment, and farming operation effect, and timely find and handle abnormal situations to prevent negative effects on crop yield or quality. At present, images in the agricultural area are usually obtained by shooting, and then analyzed, marked, etc. Based on the marked information, the crop growth, field environment, etc. Numerical information is obtained.

[0003] The marking of the image includes the marking of categories such as crops, insects, grass, etc., the number and position, the marking of the growth of the crops, and the marking of the type of each region in the image, etc. The workload of marking the image is huge. In the existing way, the image is marked by the operation personnel. First, the number of images to be processed is huge, which will require a large amount of human resources. In addition, accurate identification and annotation of the image need to rely on the professional ability of the operation personnel, which requires higher professional ability of the operation personnel. In actual application scenarios, the marking is limited, which further reduces the marking efficiency. SUMMARY

[0004] The purpose of the present application includes, for example, providing an image processing method, device, management system, electronic device and storage medium, which can improve the image annotation efficiency and guarantee the recall rate and accuracy of the final annotation information.

[0005] Embodiments of the present application can be implemented as follows:

[0006] In a first aspect, the present application provides an image processing method, which comprises:

[0007] Obtaining a to-be-labeled image of a to-be-labeled region;

[0008] Obtaining the annotation information of the to-be-labeled image based on a pre-trained annotation model, wherein the annotation model is obtained by pre-training using a sample set with annotation information;

[0009] Determining a target to-be-labeled image whose annotation information does not meet the preset requirements;

[0010] Based on the obtained correction information, the annotation information of the target to-be-labeled image is corrected to meet the preset requirements.

[0011] In an optional implementation, the method further comprises:

[0012] Adding the to-be-labeled image meeting the preset requirements to the sample set;

[0013] The labeled model is trained using the sample set to obtain a prediction model.

[0014] In an optional implementation, the annotation information includes object annotation information of the target object in the area to be annotated;

[0015] The step of obtaining the annotation information of the image to be annotated based on the pre-trained annotation model includes:

[0016] The image to be labeled is input into a pre-trained labeling model to identify the target object in the region to be labeled;

[0017] The target object is labeled to obtain the object labeling information of the target object.

[0018] In an optional implementation, the step of annotating the target object to obtain object annotation information of the target object includes:

[0019] Determine the category of the target object;

[0020] Based on the category of the target object, the target object is labeled using a corresponding labeling method;

[0021] The object annotation information of the target object is obtained based on the annotation results, and the object annotation information includes the category information of the target object.

[0022] In an optional implementation, the step of annotating the target object to obtain object annotation information of the target object includes:

[0023] The target object is defined using a bounding box.

[0024] Determine the position information of the annotation boxes and count the number of annotation boxes;

[0025] The object annotation information of the target object is obtained based on the location information and the quantity.

[0026] In an optional implementation, the annotation information includes scene annotation information for the area to be annotated;

[0027] The step of obtaining the annotation information of the image to be annotated based on the pre-trained annotation model includes:

[0028] The image to be labeled is input into a pre-trained labeling model to identify the regional state of the region to be labeled.

[0029] Based on the region status of the region to be labeled, the scene labeling information of the region to be labeled is obtained.

[0030] In an optional implementation, the image to be labeled includes multiple images within a set time period, each containing the same target object, with the target object in different states in each image;

[0031] The step of obtaining the annotation information of the image to be annotated based on the pre-trained annotation model includes:

[0032] Each image is input into a pre-trained annotation model to identify the state information of the target object in each image;

[0033] Annotations are made based on the state information of the target object in each image to obtain annotation information that represents the continuous change of the state of the target object over time.

[0034] In an optional implementation, the step of correcting the annotation information of the target image to be annotated based on the obtained correction information includes:

[0035] Identify the missing annotation information in the target image to be annotated, and supplement it using the newly added annotation information obtained from the correction information; and / or

[0036] Identify erroneous annotations in the target image to be annotated, delete the erroneous annotations, and replace them with replacement annotations from the obtained correction information; and / or

[0037] Identify erroneous annotation information in the target image to be annotated, and modify it using the modified annotation information obtained from the correction information.

[0038] In an optional implementation, the method further includes:

[0039] Obtain the image of the region to be identified;

[0040] The image to be identified is identified using a pre-trained prediction model, and the identification result corresponding to the region to be identified is obtained.

[0041] In an optional implementation, the prediction model includes multiple categories;

[0042] The step of recognizing the image to be recognized using a pre-trained prediction model includes:

[0043] The task is to obtain the recognition of the image to be recognized;

[0044] The target prediction model is determined from multiple categories of prediction models based on the identification task.

[0045] The target prediction model is used to identify the image to be identified.

[0046] In an optional implementation, the method further includes:

[0047] When the difference between the identification result and the expected result exceeds a preset range, it is determined whether to correct the identification result based on the identification result and the expected result.

[0048] If it is determined that the recognition result needs to be corrected, the image to be recognized after the recognition result is corrected is added to the sample set to train the prediction model.

[0049] In an optional implementation, the method further includes:

[0050] If it is determined that the recognition result will not be corrected, compare whether the recognition tasks of the image to be recognized and the image to be labeled belong to the same category, or compare whether the shooting states of the image to be recognized and the image to be labeled are consistent.

[0051] Feedback information is obtained based on the comparison results, and the feedback information is submitted to the management system.

[0052] Secondly, this application provides an image processing apparatus, the apparatus comprising:

[0053] The acquisition module is used to acquire the image of the region to be labeled.

[0054] The annotation module is used to obtain the annotation information of the image to be annotated based on a pre-trained annotation model, wherein the annotation model is trained in advance using a sample set with annotation information;

[0055] The determination module is used to identify target images whose annotation information does not meet preset requirements.

[0056] The correction module is used to correct the annotation information of the target image to be annotated based on the obtained correction information, so as to meet the preset requirements.

[0057] Thirdly, this application provides a management system, including:

[0058] An interaction module is used to receive tasks to be processed, including annotation tasks and / or recognition tasks, wherein the annotation task includes an image of a region to be annotated; and the recognition task includes an image to be recognized.

[0059] The processing module is used to process the image to be labeled according to the above image processing method to obtain an image to be labeled that meets preset requirements, wherein the image to be labeled that meets the preset requirements is used to train a prediction model, and / or to process the image to be recognized based on the prediction model according to the above image processing method to obtain an image recognition result.

[0060] Fourthly, this application provides an electronic device including a processor and a memory, wherein the memory stores a computer program executable by the processor, and the processor can execute the computer program to implement the image processing method described in any of the foregoing embodiments.

[0061] Fifthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the image processing method described in any of the foregoing embodiments.

[0062] The beneficial effects of the embodiments of this application include, for example:

[0063] This application provides an image labeling method, apparatus, electronic device, management system, and storage medium. It acquires an image of a region to be labeled, and obtains labeling information for the image based on a pre-trained labeling model trained using a sample set containing labeled information. It identifies target images whose labeling information does not meet preset requirements, and corrects the labeling information of these target images based on obtained correction information to meet the preset requirements. Thus, the obtained labeling model can be used to label the image, and the labeling information can be corrected further, reducing the workload of labeling personnel, improving labeling efficiency, and ensuring the recall and accuracy of the final labeled information. Attached Figure Description

[0064] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0065] Figure 1 This is one of the schematic flowcharts of the image processing method provided in the embodiments of this application;

[0066] Figure 2 for Figure 1 One of the flowcharts for the sub-steps included in step S102;

[0067] Figure 3 This is one of the target object annotation diagrams provided in the embodiments of this application;

[0068] Figure 4 This is the second schematic diagram illustrating the target object annotation provided in the embodiments of this application;

[0069] Figure 5 for Figure 1The second flowchart of the sub-steps included in step S102;

[0070] Figure 6 This is a schematic diagram of the regional status provided in an embodiment of this application;

[0071] Figure 7 for Figure 1 The third flowchart of the sub-steps included in step S102;

[0072] Figure 8 A second schematic flowchart illustrating the image processing method provided in this application embodiment;

[0073] Figure 9 The third schematic flowchart of the image processing method provided in the embodiments of this application;

[0074] Figure 10 for Figure 9 A flowchart of the sub-steps included in step S107;

[0075] Figure 11 The fourth schematic flowchart of the image processing method provided in the embodiments of this application;

[0076] Figure 12 This is a functional block diagram of the image processing method provided in the embodiments of this application;

[0077] Figure 13 A structural block diagram of the management system provided in the embodiments of this application;

[0078] Figure 14 A structural block diagram of an electronic device provided in an embodiment of this application.

[0079] Icons: 10-Image processing device; 110-Acquisition module; 120-Annotation module; 130-Determination module; 140-Correction module; 20-Electronic device; 210-Processor; 220-Memory; 230-Communication interface. Detailed Implementation

[0080] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0081] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0082] It should be noted that similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined in subsequent figures.

[0083] It should be noted that, where there is no conflict, the features in the embodiments of this application can be combined with each other.

[0084] In agricultural production management, inspecting crop growth, field environment conditions, and the effectiveness of agricultural operations helps to promptly identify and report abnormalities, which is crucial for preventing the final crop yield or quality from being affected. Therefore, in current agricultural management, farm workers at all levels obtain farmland information through on-site inspections at various stages of crop growth.

[0085] However, manual patrols have many drawbacks. For example, the amount and quality of information obtained are limited by human energy, sensory abilities, and professional skills. Patrol personnel are limited by physical strength and time constraints, limiting the area they can patrol each day. Furthermore, their field of vision is limited, and due to observation angles, some areas may be missed, making comprehensive patrols difficult, especially in large areas with limited visibility. In addition, information analysis based on observations relies on the patrol personnel's professional skills and experience; however, the professional skills of frontline farm managers often vary, leading to less professional and accurate information analysis.

[0086] To address the aforementioned shortcomings, a new approach has emerged that uses photographic equipment to capture images of farmland, which are then uploaded to a management system. The collected images are then analyzed to obtain information such as crop growth and field environmental conditions. This method allows managers to understand farmland conditions by viewing images instead of conducting on-site inspections, significantly improving the efficiency of information collection. Furthermore, image-based farmland monitoring overcomes the limitations of manual inspections, where certain areas are obscured due to angle and field-of-view constraints.

[0087] However, this highly efficient information collection method also means a large number of images need to be analyzed and processed. Managers need to annotate a large number of images, including labeling the categories, quantities, and locations of crops, insects, and grasses, as well as the growth status of crops. Based on the annotation results, they can understand the overall condition of crops and farmland. This extensive annotation work consumes significant human resources and time. Furthermore, the annotation work relies on the professional skills of managers; therefore, a lack of qualified personnel further leads to low annotation efficiency.

[0088] Therefore, to address the aforementioned technical problems, this application provides an image processing method for image annotation. An annotation model is pre-trained using a sample set containing annotation information. This model is then used to annotate the remaining images requiring annotation. Furthermore, target images whose annotation information does not meet preset requirements are identified. The annotation information of these target images is then corrected based on obtained correction information to meet the preset requirements. This approach reduces the annotation workload for operators, improves annotation efficiency, and ensures the recall and accuracy of the final annotated information.

[0089] Please see Figure 1 , Figure 1 A schematic flowchart illustrating the image processing method provided in this application embodiment, comprising the following steps:

[0090] S101, Obtain the image of the region to be labeled.

[0091] S102, based on the pre-trained annotation model, obtain the annotation information of the image to be annotated. The annotation model is obtained by pre-training using a sample set with annotation information.

[0092] S103, Identify the target image whose annotation information does not meet the preset requirements.

[0093] S104, Based on the obtained correction information, the annotation information of the target image to be annotated is corrected to meet the preset requirements.

[0094] In some possible examples, the area to be labeled can be a farmland-related area, such as an area containing farmland, the surrounding environment, and specific objects within the farmland. The obtained image to be labeled is an image containing the area to be labeled, which can be an image acquired using a sensing device. The sensing device can be a remote sensing drone or a smart camera set up next to the farmland. When using a remote sensing drone for acquisition, the image to be labeled can be obtained by segmenting video captured by a camera mounted on the drone, or it can be an image directly captured by a camera mounted on the drone.

[0095] After images are acquired, they can be automatically uploaded to a management system. This system can be used, but is not limited to, providing agricultural management services. It can be installed as a software product on a terminal or server, or it can be a combination of hardware and software. Images uploaded to the management system can be saved to an image library, from which operators or ordinary users can extract and process images.

[0096] In some possible examples, different sensing devices can be associated with different work departments. In this way, after the sensing devices collect images and upload them to the management system, the workers in the work departments associated with the sensing devices can have the right to view and annotate the corresponding images.

[0097] In some possible examples, the annotation model is pre-trained using a sample set with labeled information. The annotation model can be constructed from a neural network architecture, which may include, but is not limited to, convolutional neural networks, deep convolutional inverse graph networks, recurrent neural networks, etc.

[0098] The sample set can contain multiple sample images, each containing an area related to farmland. Each sample image can have annotation information, which can be obtained manually. In other words, a series of images requiring annotation can be divided into sample images and images to be annotated. The sample images can be pre-annotated manually to obtain annotation information.

[0099] In some possible examples, sample images can be further divided into training samples and test samples. The labeled model is trained using the training samples with annotation information. During training, a loss function can be used as a training guide until the loss function converges, at which point training stops. Specifically, the training samples are input into the labeled model for training, and the labeled model outputs the prediction information corresponding to the training samples. Based on the annotation and prediction information of the training samples, the value of the loss function can be calculated. After adjusting the model parameters of the labeled model, training continues until the loss function converges, at which point training can stop.

[0100] The accuracy and recall of the obtained annotation model can be tested using test samples. When the accuracy and recall of the annotation model reach a certain level, it indicates that the annotation model has a certain annotation capability and can be used for annotation processing of images that need to be labeled.

[0101] Since the purpose of using annotation models to annotate only a portion of the images that need annotation is to reduce the workload of manual annotation, the number of sample images with annotation information in the sample set may be relatively small. Therefore, the annotation model may have limited information about the sample images it learns from, leading to deficiencies in accuracy and recall.

[0102] Therefore, in some possible examples, based on the annotation information of the image to be annotated obtained from the annotation model, it is possible to identify target images that do not meet the preset requirements. These preset requirements may include, but are not limited to, at least one of the following: all objects in the image that need to be labeled have been labeled; the specific annotation methods for the objects (such as annotation boxes, annotation colors, annotation symbols, etc.) are correct; and objects that do not need to be labeled have not been labeled.

[0103] Accordingly, failure to meet the preset requirements can be expressed as at least one of the following: not all objects that need to be labeled in the image have been labeled (i.e., there are objects that have been omitted from the labeling), the specific labeling method of the objects is incorrect, or objects that do not need to be labeled have been labeled.

[0104] For target images that do not meet the preset requirements, corresponding correction information can be obtained. This correction information can be obtained based on the operator's actions, such as the operator inputting annotation information for omitted objects, the operator inputting modification information for correcting the annotation method of objects, or the operator performing a deletion operation to remove the annotation information of objects that do not need to be labeled. Furthermore, the annotation information of the target image can be corrected based on the correction information so that the annotation information of the target image meets the preset requirements.

[0105] It should be noted that for multiple images requiring annotation, the annotation model can accurately annotate a portion of them, while for the remaining images whose annotation information does not meet the preset requirements, only the parts requiring correction need to be corrected based on the correction information. The workload for correction is significantly reduced compared to directly annotating the entire image. Therefore, using an annotation model for annotation and then correction can greatly reduce the workload of operators, improve annotation efficiency, and ensure the accuracy and recall of the final annotation information.

[0106] In some possible examples, the sample image and the image to be labeled can be acquired by the sensing device under the same shooting conditions. Shooting conditions include factors such as shooting angle, shooting height, and lighting conditions during shooting.

[0107] Crops and farmland environments change relatively little in a short period, and the image features of images captured by sensing devices at the same shooting angle or height remain largely unchanged. Therefore, using the shooting conditions when the sample images were taken can more likely reproduce similar image features, improving the labeling accuracy of the annotation model. For example, if a labeling model for identifying weeds is trained using sample images taken by a remote sensing drone at a flight altitude of 1.5 meters, the model can learn the image features of weeds in the 1.5-meter image. However, if the image to be labeled is taken by a remote sensing drone at a flight altitude of 10 meters, many weed features that would have been visible in the 1.5-meter image will become blurred, causing the labeling model to fail to identify and label them.

[0108] Therefore, standardizing image acquisition through the above methods helps ensure the accuracy of image annotation.

[0109] The agricultural information required may differ at different stages of agricultural activity. For example, during the crop growth stage, information about the crop itself is needed, while after harvest, information about the farmland may be required. Therefore, image annotation may involve different annotation tasks; some require labeling regions, while others require labeling objects contained within those regions. Thus, to enable annotation models to focus on different annotation tasks, separate annotation models capable of performing different tasks can be pre-trained.

[0110] In some possible examples, if the annotation task requires labeling specific objects within a region, the annotation information may include the object annotation information of the target objects in the region to be labeled. Please refer to [link to relevant documentation]. Figure 2 One possible implementation of step S102 above is as follows:

[0111] S1021A: Input the image to be labeled into the pre-trained labeling model to identify the target object in the region to be labeled.

[0112] S1022A, annotate the target object to obtain the object annotation information of the target object.

[0113] In some possible examples, the target object could be crops, grass, insects, burrows, or other objects in a field. An object detection and annotation model capable of annotating such target objects can be pre-trained.

[0114] Furthermore, the target object can also be a specific area within farmland, such as a high mound area, a low-lying area, or an area lacking seedlings. An image segmentation and annotation model capable of annotating such target objects can be pre-trained.

[0115] The above-mentioned annotation models can identify target objects in an image and then annotate them.

[0116] In some possible examples, when annotating a target object, the target object can be defined using annotation boxes, the position information of the annotation boxes can be determined, and the number of annotation boxes can be counted. Based on the position information and the number, the object annotation information of the target object can be obtained.

[0117] The annotation box can be the smallest box that can enclose the target object, and its shape can be a rectangle, circle, rhombus, or other irregular shape. For example, Figure 3 As shown, the label box can be a rectangle that defines the wheat seedling. For example, as... Figure 4 As shown, the label box can be an irregularly shaped box that defines the area in the farmland where seedlings are missing.

[0118] In some possible examples, the position of the geometric center point of the annotation box can be used as the position of the annotation box. This position is the relative position of the annotation box in the image to be annotated. It can be combined with the absolute position of the image in the world coordinate system when the sensing device captured the image to be annotated, so as to determine the absolute position of each annotation box, and then determine the absolute position of the target object enclosed by the annotation box.

[0119] By determining the position information of the annotation boxes and counting the number of annotation boxes, the object annotation information of the target objects obtained can include the position information of each target object and the number of target objects.

[0120] In some possible examples, when the target objects in the image to be labeled are such as high mound areas, low-lying areas, or areas lacking seedlings, after defining each type of area with a label box, the area proportion of the same type of area in the image to be labeled, or the area proportion of the farmland area in the image to be labeled, can also be calculated.

[0121] Thus, the obtained area proportions can provide data support for intelligent farmland management. For example, the obtained area proportions of areas with missing seedlings can provide a basis for deciding whether to replant seedlings in the farmland.

[0122] Alternatively, in order to distinguish specific target objects, such as whether they are crops or weeds, the labeling of target objects can be achieved in the following ways in some possible examples:

[0123] The category of the target object is determined. Based on the category of the target object, the target object is labeled using the corresponding labeling method. The object labeling information of the target object is obtained according to the labeling results. This object labeling information includes the category information of the target object.

[0124] In some possible examples, different labeling methods can be used for different categories of target objects. This could involve using different colors, such as different colored borders, or different shapes, such as different shaped borders or symbols. For instance, crops could be labeled with rectangles, while weeds could be labeled with diamonds. Similarly, burrows in farmland could be labeled with red borders, while pests in farmland could be labeled with yellow borders.

[0125] By using different annotation methods to annotate different categories of target objects as described above, we can obtain the object annotation information of the target objects, which can contain the category information of the target objects.

[0126] In this way, when conducting subsequent analysis of farmland-related conditions based on object annotation information, targeted analysis and processing can be carried out from various category perspectives.

[0127] In some possible examples, if the annotation task requires labeling farmland-related information, the annotation information may include scene annotation information for the area to be labeled. Please refer to [link to relevant documentation]. Figure 5 One possible implementation of step S102 above is as follows:

[0128] S1021B inputs the image to be labeled into a pre-trained labeling model to identify the regional state of the region to be labeled.

[0129] S1022B: Based on the region status of the region to be labeled, obtain the scene labeling information of the region to be labeled.

[0130] In some possible examples, the state of the region to be labeled can include, for example, the state of farmland such as drying or water retention. Based on this, an image classification and labeling model capable of labeling region states can be pre-trained.

[0131] These types of annotation tasks do not require labeling the location, quantity, or other information of objects in an image; they only require identifying the state of the region and then labeling it. Annotation can be done using text or symbols.

[0132] In some possible examples, the annotation model can determine the region status of the area to be annotated by identifying the pixel features of the region in the image. For example, the image to be annotated could be an RGB image, which obtains various colors by varying and superimposing the red, green, and blue color channels.

[0133] When farmland within the area to be labeled is in a drying state, the color of the farmland area should lean towards yellow, such as...Figure 6 As shown on the left, when farmland is in a water-retaining state, the color of the farmland area should lean towards green, white, and blue, such as... Figure 6 As shown on the right side of the image. Therefore, the annotation model can identify the pixel features of farmland areas, obtain the RGB values ​​of farmland areas, determine the color information of farmland areas based on the RGB values, and then determine the regional state of farmland areas based on the color information.

[0134] By identifying and labeling the regional status of farmland areas, it is possible to determine whether farmland within a certain range is in a water-retaining or drying state, thereby enabling scientific management of farmland.

[0135] In some possible examples, if it is necessary to obtain specific growth information about crops, such as leaf age and number of tillers, it is not only necessary to determine the location of the crops but also to analyze their specific state. Furthermore, since it may not be possible to label every stage of crop growth during the annotation phase, some features not learned during the annotation phase may cause the model to have difficulty accurately identifying them if they appear in subsequent stages requiring actual recognition.

[0136] Based on this consideration, in another possible implementation, the obtained image to be labeled can contain multiple images within a set time period, each containing the same target object, with the target object in different states in each image. Please refer to [link to relevant documentation]. Figure 7 The above step S102 can be achieved in the following way:

[0137] S1021C inputs each image into a pre-trained annotation model to identify the state information of the target object in each image.

[0138] S1022C, based on the state information of the target object in each image, annotates to obtain annotation information that characterizes the continuous change of the state of the target object over time.

[0139] In some possible examples, multiple images within a given time period may contain the same target object, such as wheat seedlings or fruit trees. However, the target object may be in different states in each image, for example, wheat seedlings may be at different growth stages. An image regression annotation model capable of state annotation for such target objects can be pre-trained.

[0140] Because the annotation model is pre-trained, it can analyze each image and identify the state information of the target objects within them. The state information of the target objects could be, for example, the leaf state features of wheat seedlings.

[0141] Based on the state information of the target object in each image, annotation information can be determined. This annotation information can be numerical, symbolic, etc. For example, the leaf age of wheat seedlings in each image can be numerically labeled.

[0142] Since the target object is at different stages in each image, the overall annotation information can reflect the continuous change of the target object's state over time. For example, the leaf age of rice at each stage from the early growth stage to heading. If the annotation information is numerical, it can be numerical or curvilinear information showing a linear trend in the target object's state over a certain period of time. In short, it can be understood as the numerical value of the target object's state continuously changing over a certain period of time with a linear trend.

[0143] While labeling images, the annotation model learns a series of annotation information showing that the state of target objects in the images changes linearly over time. Subsequently, when using a prediction model to identify images, even if the state of the target object in the image was not present during the model learning phase, the model can still predict the state information of the target object based on the learned linearly changing information, ensuring the accuracy of subsequent identification.

[0144] For example, for a specific target object, the annotation model learns and annotates the feature information of the target object at leaf ages of 1, 3, 5, 7, 9, and 11. Based on this series of feature information, linearly varying annotation information of the target object can be obtained. If the leaf age of the target object in the image to be identified is 2, although the annotation model has not learned the feature information of the target object at leaf age 2, based on the linearly varying annotation information, the feature information of the target object at leaf age 2 can be imported into this linearly varying annotation information for comparison. This allows it to be determined that the leaf age of the target object in the image to be identified is between 1 and 3 leaf ages. In this way, the leaf age value of the target object in the image to be identified can be determined.

[0145] By pre-constructing annotation models that can be used to perform different annotation tasks, including object detection annotation models, image segmentation annotation models, image classification annotation models, and image regression annotation models, the images to be annotated can be specifically labeled to obtain annotation information. This reduces the annotation workload for operators.

[0146] However, since the number of sample images used for training the model is relatively small, the annotation information obtained from the model may not meet some of the preset requirements.

[0147] In some possible examples, for target images whose annotation information does not meet the preset requirements, the annotation information can be corrected based on the correction information provided by the operators.

[0148] When correcting annotation information, one possible approach is to identify the missing annotation information in the target image to be annotated and supplement it using the newly added annotation information obtained from the correction information.

[0149] For example, if the annotation task requires identifying and framing wheat seedlings in an image, but the annotation model outputs annotation information that does not include all wheat seedlings within the bounded boxes, then these unframed wheat seedlings can be identified, and bounded boxes can be drawn to supplement the annotation.

[0150] As another possible implementation, erroneous annotation information in the target image to be annotated can be identified, the erroneous annotation information can be deleted, and the replacement annotation information in the obtained correction information can be used to replace it.

[0151] For example, if the annotation task requires defining and labeling wheat seedlings and weeds in an image, and wheat seedlings and weeds belong to different categories, then different annotation methods need to be used. For instance, if the annotation model incorrectly labels some wheat seedlings using the annotation method corresponding to weeds, then the incorrectly labeled wheat seedlings can be deleted, and the correct annotation method for wheat seedlings can be used for replacement.

[0152] As another possible approach, erroneous annotation information in the target image to be annotated can be identified, and the obtained correction information can be used to modify the annotation information.

[0153] For example, if the annotation task requires defining and labeling missing seedling areas in an image, this type of task often involves drawing boundaries along the edges of these areas. If the annotation model outputs annotation boxes with some misaligned edges, then the misaligned parts can be modified and adjusted based on the obtained modified annotation information.

[0154] It should be noted that when correcting the annotation information, any one of the above methods can be used, or a combination of the above methods can be used for comprehensive correction, depending on the specific needs.

[0155] In some possible examples, the labeled images can be used for subsequent comprehensive analysis of farmland conditions. In this case, the labeled images that do not meet the preset requirements should be corrected as much as possible to ensure the accuracy of the labeled information in the corrected images, so as to provide a valid basis for comprehensive analysis.

[0156] In some possible examples, the images to be labeled can be used to further train the labeling model after labeling. In this case, before correcting the labeling information of the images to be labeled, images whose labeling information differs significantly from the actual required labeling results can be selected. The significant difference between the labeling information and the actual required labeling results for these images indicates that correcting them would require considerable effort, perhaps even exceeding the effort required for manual labeling. Therefore, these images can be extracted.

[0157] However, due to insufficient feature learning in these extracted images, the annotation information of these images often differs significantly from the actual required annotation results. Therefore, to improve the annotation performance of the annotation model, the extracted images can be manually annotated to obtain annotation information. This annotated images can then be used to further train the annotation model, enabling it to learn features from these images.

[0158] For images whose annotation information does not meet the preset requirements but is not significantly different from the actual required annotation results, these images can be corrected based on the correction information. The annotation model can then be trained again based on the corrected images.

[0159] Therefore, please refer to Figure 8 In some possible examples, the image processing method provided in this embodiment may further include the following steps:

[0160] S105, add the images to be labeled that meet the preset requirements to the sample set.

[0161] S106, the labeled model is trained using the sample set to obtain the prediction model.

[0162] Among them, the image to be labeled that meets the preset requirements is an image that has been labeled with information. It can include the image to be labeled whose labeling information directly output by the labeling model meets the preset requirements, and it can also include the image to be labeled whose labeling information output by the labeling model does not meet the preset requirements, but whose labeling information meets the preset requirements after correction.

[0163] Using the images to be labeled and sample images from the sample set, the labeling model is further trained. The constructed loss function is used as a training guide until the loss function converges, resulting in a trained prediction model. This prediction model can then be used in the application phase for image recognition.

[0164] Please see Figure 9 In some possible examples, the image processing method provided in this embodiment may further include the following steps:

[0165] S107, Obtain the image of the region to be identified.

[0166] S108, the prediction model is used to identify the image to be identified, and the identification result corresponding to the region to be identified is obtained.

[0167] During the application phase, operators can remotely control smart cameras positioned near farmland or use remote sensing drones to capture images of the area to be identified. To ensure the accuracy of the predictive model's identification, the images to be identified and the aforementioned images to be labeled can be captured under the same shooting conditions. Shooting conditions include shooting angle, shooting height, and lighting conditions during shooting.

[0168] To simplify the shooting state settings for different recognition tasks, different task templates can be created based on the shooting state when pre-collecting images for model training. Each task template can include information such as the location of the point of interest, shooting height, shooting angle, lighting conditions, and the recognition task. Thus, when performing a recognition task for a specific point of interest later, the task template for that point of interest can be called, and the task parameters can be set based on the information in the task template.

[0169] It should be noted that the image to be identified in this embodiment can also be an image obtained through other channels, such as an image directly downloaded from the Internet or an image copied from a storage device.

[0170] The prediction model is used to identify the image to be identified. The identification results may include the location and quantity of crops, weeds, pests, holes, etc., as well as information such as the leaf age of crops or the status information of farmland areas.

[0171] It is evident that different recognition tasks require the identification of different objects. As described above, annotation models can include various categories, which can be used to annotate specific target objects, region states, and target object states, respectively. Correspondingly, prediction models obtained by further training based on prediction models can also include various categories.

[0172] Please see Figure 10In order to perform targeted recognition processing on the image to be recognized, one possible implementation of step S107 is as follows:

[0173] S1071, Obtain the recognition task of the image to be recognized.

[0174] S1072, determine the target prediction model from multiple categories of prediction models based on the identification task.

[0175] S1073 uses a target prediction model to identify the image to be identified.

[0176] The recognition task of the image to be recognized can be the recognition of a specific target object, the recognition of the state of a region, or the recognition of the state of a target object. Under different recognition tasks, the information contained in the acquired image to be recognized may differ. For example, when the recognition task is the recognition of a specific target object, the image to be recognized may contain objects such as crops, weeds, pests, and burrows. When the recognition task is the recognition of the state of a region, the image to be recognized may not contain specific objects and may only contain an image of farmland.

[0177] Therefore, the information contained in the images to be identified differs under different recognition tasks. Accordingly, in order to accurately identify the information in a targeted manner, prediction models corresponding to different categories can be adopted.

[0178] After obtaining the recognition results of the image to be recognized, the results can be displayed. For example, they can be displayed through the management system's interface or sent to the terminal device held by the operator.

[0179] To further improve the recognition accuracy of the prediction model, the recognition results of the image to be recognized can be used to determine whether further training of the prediction model using the image to be recognized is necessary. Please refer to [link / reference]. Figure 11 Therefore, the image processing method provided in this embodiment may further include the following steps:

[0180] S109, when the difference between the recognition result and the expected result exceeds a preset range, determine whether to correct the recognition result based on the recognition result and the expected result. If it is determined that the recognition result should be corrected, perform the following step S110.

[0181] S110, add the image to be identified after the recognition result is corrected to the sample set in order to train the prediction model.

[0182] The expected result can be the actual result of the image to be recognized. When determining whether the difference between the recognition result and the expected result is within a preset range, if the recognition task is to identify a specific object, this can be done by judging whether the difference between the number of specific objects identified in the recognition result and the actual number of specific objects in the expected result is within a preset range. Alternatively, if the recognition task is to identify the state of a specific object, this can be done by judging whether the difference between the state value of the specific object in the recognition result and the state value of the specific object in the expected result is within a preset range. Or, if the recognition task is to recognize and draw a certain region, this can be done by judging whether the portion of the drawing information in the recognition result that matches the drawing information in the expected result exceeds a certain percentage.

[0183] When the difference between the recognition result and the expected result exceeds the preset range, it indicates that the recognition accuracy or recall rate of the prediction model for the image to be recognized is low. In this case, it is advisable to use the image to be recognized to further train the prediction model in order to improve the recognition accuracy of the prediction model for such images in the future.

[0184] However, the large discrepancy between the recognition result and the expected result may be due to poor model recognition performance, mismatch between the selected model category, or the image itself.

[0185] If the poor recognition performance is due to the model's inability to recognize images, then although the difference between the predicted model's recognition result and the expected result exceeds the preset range, there should still be some consistency between the two. That is, the predictive model can achieve partial accurate recognition of the image to be recognized. Further training the predictive model using such images can improve its recognition performance. Therefore, in this case, it can be determined that the recognition result of the image to be recognized can be corrected, and the corrected image can be added to the sample set to train the predictive model.

[0186] Optionally, the method for correcting the recognition result of the image to be recognized can be the same as the method for correcting the annotation information of the image to be labeled, which will not be elaborated here.

[0187] However, if the discrepancy between the prediction model's recognition result and the expected result is due to the image itself, it may not only exceed the preset range, but the two may also be completely different or only partially consistent. In this case, the prediction model used may not be suitable for processing this type of image to be recognized, meaning that the image to be recognized does not help in further training the prediction model.

[0188] Therefore, in some possible examples, if the difference between the recognition result and the expected result exceeds a preset range, it can be determined whether the recognition result should be corrected based on the recognition result and the expected result. If it is determined that the recognition result should be corrected, it indicates that such images to be recognized can be used to further train the prediction model. The images to be recognized with corrected recognition results can be added to the sample set to train the prediction model.

[0189] Please refer to 11. If it is determined that the recognition result will not be corrected, then perform the following steps S111 and S112.

[0190] S111, compare whether the recognition tasks of the image to be recognized and the image to be labeled belong to the same category, or compare whether the shooting states of the image to be recognized and the image to be labeled are consistent.

[0191] S112. Obtain feedback information based on the comparison results and submit the feedback information to the management system.

[0192] If it is determined that no correction is made to the recognition results, it indicates that this type of image to be recognized is not conducive to further training of the prediction model. The prediction model is trained using the images to be labeled. In order to determine why this type of image to be recognized is not conducive to further training of the prediction model, we can compare whether the recognition tasks of the images to be recognized and the images to be labeled belong to the same category, or compare whether the shooting states of the images to be recognized and the images to be labeled are consistent. Feedback information is obtained based on the comparison results and submitted to the management system.

[0193] In some possible examples, the information contained in the image may differ depending on the recognition task. For example, when recognizing a specific target object, the image usually contains the target object, while when recognizing the state of a region, the image contains information such as the land surface.

[0194] When the image to be labeled is for a specific recognition task, it contains image information corresponding to that task. The prediction model can learn the feature information of the image for that recognition task. If the recognition tasks of the image to be identified and the image to be labeled belong to the same category, the image to be identified also contains image information corresponding to that category of recognition tasks. Since the prediction model has already learned the image features for that category of recognition tasks, it can be used to identify the image to be identified.

[0195] However, if the recognition tasks of the image to be recognized and the image to be labeled do not belong to the same category—for example, the image to be labeled is for labeling a specific target object, while the image to be recognized is for recognizing a region's state—then the image to be recognized does not contain the same image information as the image to be labeled. Since the prediction model has not learned the features of the image information in the image to be recognized, it cannot be used to recognize the image, resulting in a large discrepancy between the recognition result and the expected result, and failing to obtain a result relevant to the expectation.

[0196] By comparing the recognition tasks of the image to be recognized and the image to be labeled, we can obtain recognition results that characterize whether the image to be recognized and the image to be labeled have the same image information.

[0197] In some possible examples, if the shooting conditions of the image to be identified and the image to be labeled are inconsistent, including any one or more of the shooting angle, shooting height, and lighting conditions, then the image information in the image to be identified and the image to be labeled will be inconsistent.

[0198] For example, if the image to be labeled is taken at a height of 1.5 meters, the prediction model will learn the image features at that height. However, if the image to be identified is taken at a height of 10 meters, the information from a 10-meter image differs significantly from that from a 1.5-meter image; some objects that were originally clearly visible will become blurry. The resulting prediction model will struggle to accurately identify the image. Furthermore, because prediction models also have training specifications, they are not designed to learn from images taken in all conditions. Therefore, such images are not conducive to further training of the prediction model.

[0199] By comparing the results obtained in the above manner, the operators can submit them to the management system. This allows them to understand why such images are not conducive to the further learning of the prediction model, and thus make adjustments to the subsequent processing.

[0200] Please see Figure 12 This application also provides an image processing device 10, which includes an acquisition module 110, an annotation module 120, a determination module 130, and a correction module 140.

[0201] The acquisition module 110 is used to acquire the image of the region to be labeled.

[0202] The annotation module 120 is used to obtain the annotation information of the image to be annotated based on the annotation model obtained in advance. The annotation model is obtained in advance by training a sample set with annotation information.

[0203] The determination module 130 is used to determine the target image to be labeled if the annotation information does not meet the preset requirements.

[0204] The correction module 140 is used to correct the annotation information of the target image to be annotated based on the obtained correction information, so as to meet the preset requirements.

[0205] It is understandable that the acquisition module 110, the annotation module 120, the determination module 130 and the correction module 140 can be used to perform steps S101 to S104 to achieve the corresponding technical effects.

[0206] Optionally, the image processing apparatus 10 may further include a training module, which can be used for:

[0207] Add the images to be labeled that meet the preset requirements to the sample set;

[0208] The labeled model is trained using the sample set to obtain the prediction model.

[0209] Optionally, the annotation information includes object annotation information of the target objects in the area to be annotated, and the annotation module 120 can be used for:

[0210] The image to be labeled is input into a pre-trained labeling model to identify the target object in the region to be labeled.

[0211] The target object is annotated to obtain the object annotation information of the target object.

[0212] Optionally, the annotation module 120 can be used for:

[0213] Determine the category of the target object;

[0214] Based on the category of the target object, the target object is labeled using the corresponding annotation method;

[0215] The object annotation information of the target object is obtained based on the annotation results. The object annotation information includes the category information of the target object.

[0216] Optionally, the annotation module 120 can be used for:

[0217] Use the annotation box to define the target object;

[0218] Determine the location information of the annotation boxes and count the number of annotation boxes;

[0219] Based on location information and quantity, obtain the object annotation information of the target object.

[0220] Optionally, the annotation information includes scene annotation information for the area to be annotated, and the annotation module 120 can be used for:

[0221] The image to be labeled is input into a pre-trained labeling model to identify the region status of the region to be labeled.

[0222] Based on the region status of the region to be labeled, the scene labeling information of the region to be labeled is obtained.

[0223] Optionally, the image to be labeled includes multiple images within a set time period, each containing the same target object, with the target object in different states in each image. The labeling module 120 can be used for:

[0224] Each image is input into a pre-trained annotation model to identify the state information of the target object in each image;

[0225] Annotations are made based on the state information of the target object in each image to obtain annotation information that represents the continuous change of the state of the target object over time.

[0226] Optionally, the above-mentioned correction module 140 can be used for:

[0227] Identify missing annotation information in the target image to be annotated, and supplement it using the newly added annotation information obtained from the correction information; and / or

[0228] Identify erroneous annotations in the target image to be annotated, delete the erroneous annotations, and replace them with replacement annotations from the obtained correction information; and / or

[0229] Identify erroneous annotations in the target image to be annotated, and then modify them using the corrected annotations obtained from the correction information.

[0230] Optionally, the image processing apparatus 10 may further include a recognition module, which can be used for:

[0231] Obtain the image of the region to be identified;

[0232] The pre-trained prediction model is used to identify the image to be identified, and the identification result corresponding to the region to be identified is obtained.

[0233] Optionally, the image to be identified and the image to be labeled are captured under the same shooting conditions.

[0234] Optionally, the prediction model includes multiple categories, and the above-mentioned identification module can be used for:

[0235] The task is to acquire the image to be recognized.

[0236] The target prediction model is determined from multiple categories of prediction models based on the identification task;

[0237] The target prediction model is used to identify the image to be identified.

[0238] Optionally, the image processing apparatus 10 may further include a determination module, which may be used for:

[0239] When the difference between the identification result and the expected result exceeds a preset range, a determination is made based on the identification result and the expected result to determine whether to correct the identification result.

[0240] If it is determined that the recognition result needs to be corrected, the image to be recognized after the correction is added to the sample set to train the prediction model.

[0241] Optionally, the above-mentioned judgment module can also be used for:

[0242] If it is determined that no correction will be made to the recognition results, compare whether the recognition tasks of the image to be recognized and the image to be labeled belong to the same category, or compare whether the shooting states of the image to be recognized and the image to be labeled are consistent.

[0243] Based on the comparison results, feedback information is obtained and submitted to the management system.

[0244] Please see Figure 13 This application also provides a management system, which includes an interactive module and a processing module interconnected with each other. The interactive module and the processing module can be installed as software products in a terminal or server, or they can be a combination of hardware and software. The interactive module and the processing module can be electrically connected or communicatively connected; for example, they can be connected via a cable or via wireless communication methods such as Bluetooth or Wi-Fi.

[0245] The interaction module provides human-computer interaction functions, such as input / output devices like keyboards, mice, and touch controls. Alternatively, the interaction module can be a software function module underlying the hardware for this input / output. The processing module can be a processor, such as in a terminal or server, or a software function module that can be called and executed by the processor.

[0246] When users need to perform image processing, such as recognizing a region of an image, they often lack effective prediction models for processing these images. Therefore, it is necessary to first annotate the images and then use the annotated images to train a high-performance prediction model.

[0247] Based on this, in this embodiment, the interactive device can be used to receive tasks to be processed, including annotation tasks and / or recognition tasks, wherein the annotation task includes an image of a region to be annotated; and the recognition task includes an image to be recognized.

[0248] However, relying entirely on manual image annotation would result in low efficiency and both low recall and accuracy. Therefore, in this embodiment, a subset of images can be pre-annotated manually to form a sample set with annotation information, and the annotation model can be initially trained using this sample set.

[0249] Based on this, in this embodiment, the processing module can be used to obtain the annotation information of the image to be annotated based on the pre-trained annotation model, then determine the target image to be annotated whose annotation information does not meet the preset requirements, and correct the annotation information of the target image to be annotated based on the obtained correction information to meet the preset requirements.

[0250] In this way, by using the annotation model to obtain the annotation information of the images to be annotated, and then correcting the annotation information of the target images that do not meet the preset requirements, the workload of the annotators can be reduced, the annotation efficiency can be improved, and the recall and accuracy of the final annotation information can be guaranteed.

[0251] Based on this, the processing module can also be used to obtain the image of the region to be identified in the recognition task, and use the prediction model to identify the image to obtain the recognition result corresponding to the region to be identified. The prediction model is obtained by training the annotation model using a sample set of images to be labeled that meet the preset requirements, where the images to be labeled that meet the preset requirements are images that have already been annotated.

[0252] By adding images that meet preset requirements to the sample set, including images directly labeled by the annotation model and images labeled by the annotation model and corrected using correction information, the annotation model is further trained to obtain a prediction model. The well-performing prediction model obtained through training can effectively identify the region to be identified and output the identification result.

[0253] It should be noted that the management system provided in this embodiment can implement any of the image processing methods in the foregoing embodiments. For any details not covered in this embodiment, please refer to the foregoing embodiments.

[0254] This application also provides an electronic device, such as... Figure 14 , Figure 14This is a structural block diagram of an electronic device 20 according to an embodiment of this application. The electronic device 20 includes a communication interface 230, a processor 210, and a memory 220. The processor 210, memory 220, and communication interface 230 are electrically connected directly or indirectly to achieve data transmission or interaction. For example, these components can be electrically connected to each other through one or more communication buses or signal lines. The memory 220 can be used to store software programs and modules, such as the program instructions / modules corresponding to the image processing method provided in this embodiment. The processor 210 executes the software programs and modules stored in the memory 220 to perform various functional applications and data processing. The communication interface 230 can be used for signaling or data communication with other node devices. In this application, the electronic device 20 may have multiple communication interfaces 230.

[0255] The memory 220 may be, but is not limited to, random access memory (RAM), read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), etc.

[0256] Processor 210 can be an integrated circuit chip with signal processing capabilities. Processor 210 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0257] This application provides a storage medium storing a computer program thereon. When executed by processor 210, the computer program implements the image processing method as described in any of the foregoing embodiments. The computer-readable storage medium may be, but is not limited to, various media capable of storing program code, such as a USB flash drive, portable hard drive, ROM, RAM, PROM, EPROM, EEPROM, magnetic disk, or optical disk.

[0258] In summary, the image processing method, apparatus, management system, electronic device 20, and storage medium provided in this application acquire an image of a region to be labeled, and obtain labeling information for the image based on a pre-trained labeling model. This labeling model is trained in advance using a sample set containing labeled information. Target images whose labeling information does not meet preset requirements are identified, and their labeling information is corrected based on obtained correction information to meet the preset requirements. Thus, the obtained labeling model can be used to label the image, and further correction of the labeling information can reduce the workload of personnel, improve labeling efficiency, and ensure the recall and accuracy of the final labeled information.

[0259] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An image processing method, characterized in that, The method includes: Obtain the image of the region to be labeled; Based on the pre-trained annotation model, the annotation information of the image to be annotated is obtained. The annotation model is obtained by pre-training using a sample set with annotation information. Identify target images whose annotation information does not meet preset requirements; The annotation information of the target image to be annotated is corrected based on the obtained correction information to meet the preset requirements; The images to be labeled include multiple images, each containing the same target object, taken within a set time period. The target object in each image is in a different state, representing different growth stages of the target object. The labeling information includes annotations characterizing the continuous change of the target object's state over time, obtained through the following methods: Each image is input into a pre-trained annotation model to identify the state information of the target object in each image. Based on the state information of the target object in each image, annotation information is obtained to represent the continuous change of the state of the target object over time. The method further includes: A region to be identified is obtained as an image to be identified. The image to be identified is then identified using a trained prediction model to obtain a recognition result corresponding to the region to be identified. The recognition result includes the state of the target object in the image to be identified. The prediction model is obtained by training the annotation model on a sample set of images to be labeled that meet preset requirements.

2. The image processing method according to claim 1, characterized in that, The annotation information includes object annotation information of the target objects in the area to be annotated; The step of obtaining the annotation information of the image to be annotated based on the pre-trained annotation model includes: The image to be labeled is input into a pre-trained labeling model to identify the target object in the region to be labeled; The target object is labeled to obtain the object labeling information of the target object.

3. The image processing method according to claim 2, characterized in that, The step of annotating the target object to obtain the object annotation information of the target object includes: Determine the category of the target object; Based on the category of the target object, the target object is labeled using a corresponding labeling method; The object annotation information of the target object is obtained based on the annotation results, and the object annotation information includes the category information of the target object.

4. The image processing method according to claim 2, characterized in that, The step of annotating the target object to obtain the object annotation information of the target object includes: The target object is defined using a bounding box. Determine the position information of the annotation boxes and count the number of annotation boxes; The object annotation information of the target object is obtained based on the location information and the quantity.

5. The image processing method according to claim 1, characterized in that, The annotation information includes scene annotation information for the area to be annotated; The step of obtaining the annotation information of the image to be annotated based on the pre-trained annotation model includes: The image to be labeled is input into a pre-trained labeling model to identify the regional state of the region to be labeled. Based on the region status of the region to be labeled, the scene labeling information of the region to be labeled is obtained.

6. The image processing method according to claim 1, characterized in that, The step of correcting the annotation information of the target image to be annotated based on the obtained correction information includes: Identify the missing annotation information in the target image to be annotated, and supplement it using the newly added annotation information obtained from the correction information; and / or Identify erroneous annotations in the target image to be annotated, delete the erroneous annotations, and replace them with replacement annotations from the obtained correction information; and / or Identify erroneous annotation information in the target image to be annotated, and modify it using the modified annotation information obtained from the correction information.

7. The image processing method according to claim 1, characterized in that, The prediction models include multiple categories; The steps of recognizing the image to be recognized using the prediction model include: The task is to obtain the recognition of the image to be recognized; The target prediction model is determined from multiple categories of prediction models based on the identification task. The target prediction model is used to identify the image to be identified.

8. The image processing method according to claim 1, characterized in that, The method further includes: When the difference between the identification result and the expected result exceeds a preset range, it is determined whether to correct the identification result based on the identification result and the expected result. If it is determined that the recognition result needs to be corrected, the image to be recognized after the recognition result is corrected is added to the sample set to train the prediction model.

9. The image processing method according to claim 8, characterized in that, The method further includes: If it is determined that the recognition result will not be corrected, compare whether the recognition tasks of the image to be recognized and the image to be labeled belong to the same category, or compare whether the shooting states of the image to be recognized and the image to be labeled are consistent. Feedback information is obtained based on the comparison results, and the feedback information is submitted to the management system.

10. An image processing apparatus, characterized in that, The apparatus for implementing the image processing method according to any one of claims 1-9, comprising: The acquisition module is used to acquire the image of the region to be labeled. The annotation module is used to obtain the annotation information of the image to be annotated based on a pre-trained annotation model, wherein the annotation model is trained in advance using a sample set with annotation information; The determination module is used to identify target images whose annotation information does not meet preset requirements. The correction module is used to correct the annotation information of the target image to be annotated based on the obtained correction information, so as to meet the preset requirements.

11. A management system, characterized in that, include: An interaction module is used to receive tasks to be processed, including annotation tasks and / or recognition tasks, wherein the annotation task includes an image of a region to be annotated; and the recognition task includes an image to be recognized. The processing module is configured to process the image to be labeled according to any one of claims 1 to 6 to obtain an image to be labeled that meets preset requirements, wherein the image to be labeled that meets preset requirements is used to train a prediction model, and / or the image processing method according to any one of claims 7 to 9 processes the image to be recognized based on the prediction model to obtain an image recognition result.

12. An electronic device, characterized in that, It includes a processor and a memory, the memory storing a computer program executable by the processor, the processor being able to execute the computer program to implement the image processing method according to any one of claims 1-9.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the image processing method according to any one of claims 1-9.

Images