An intelligent generation method and device for a remote sensing image label frame oriented to on-orbit computing

Abstract

The application relates to the field of remote sensing image processing, and discloses a remote sensing image labeling frame intelligent generation method and device for on-orbit computing, the method comprising the following steps: acquiring a target remote sensing image set to be labeled, performing preliminary labeling on the target remote sensing image set to obtain a plurality of initial labeling frames; acquiring target remote sensing images and coordinate information corresponding to any initial labeling frame and inputting the target remote sensing images and the coordinate information into a labeling frame inspection model, the labeling frame inspection model being obtained by pre-training based on a preference data set, and any preference data sample of the preference data set comprising at least positive preference information and negative preference information; determining whether the initial labeling frame is an alignment defect labeling frame by using the labeling frame inspection model, if the initial labeling frame is determined to be an alignment defect labeling frame, re-labeling a target object corresponding to the initial labeling frame, and taking a labeling frame obtained by re-labeling as a target labeling frame of the target object. The technical scheme provided by the application can improve the generation quality and inspection efficiency of remote sensing image labeling frames.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of remote sensing image processing, and in particular to a remote sensing image label box intelligent generation method and device for on-orbit computing. BACKGROUND

[0002] With the continuous development of on-orbit computing technology, in the training process of a remote sensing image target detection model for on-orbit computing, the demand for high-quality labeling of a remote sensing image dataset used for model training is increasing. The labeling task usually involves identifying and locating target objects in an image and drawing a label box to frame the target objects.

[0003] In practical applications, the current labeling of a remote sensing image dataset usually adopts manual labeling or automatic tool labeling. Since the label box and the actual region of the target object are prone to be misaligned, the alignment quality of the label box needs to be checked and optimized. However, manual checking of the label box is inefficient and cannot meet the production needs of large-scale and high-quality datasets.

[0004] Therefore, how to efficiently identify and correct label boxes with low alignment quality to improve the generation quality and inspection efficiency of remote sensing image label boxes has become a key problem to improve the labeling quality of remote sensing image datasets and the performance of remote sensing image target detection models for on-orbit computing. SUMMARY

[0005] The present application provides a remote sensing image label box intelligent generation method and device for on-orbit computing, which can improve the generation quality and inspection efficiency of remote sensing image label boxes.

[0006] The first aspect of the present application provides a remote sensing image label box intelligent generation method for on-orbit computing, the method comprising: obtaining a target remote sensing image set to be labeled, and preliminarily labeling the target remote sensing image set to obtain a plurality of initial label boxes, wherein the initial label boxes are used to label target objects in any target remote sensing image; obtaining a target remote sensing image and coordinate information corresponding to any initial label box, and inputting the target remote sensing image and the coordinate information into a label box inspection model, wherein the label box inspection model is obtained by pre-training based on a preference dataset, and any preference data sample of the preference dataset includes at least positive preference information and negative preference information; determining whether the initial label box belongs to an alignment defect label box by using the label box inspection model, if it is determined that the initial label box belongs to an alignment defect label box, re-labeling the target object corresponding to the initial label box, and taking the re-labeled label box as a target label box of the target object.

[0007] In an embodiment, determining, by the label box inspection model, whether the initial label box belongs to the alignment defect label box includes: generating, by the label box inspection model, an alignment quality score of the initial label box, determining whether the alignment quality score is less than a preset alignment threshold; if the alignment quality score is less than the preset alignment threshold, determining that the initial label box belongs to the alignment defect label box, and if the alignment quality score is greater than or equal to the preset alignment threshold, determining that the initial label box does not belong to the alignment defect label box.

[0008] In an embodiment, if it is determined that the initial label box does not belong to the alignment defect label box, the initial label box is taken as a target label box of the corresponding target object.

[0009] In an embodiment, the initial label box includes a horizontal label box and a rotated label box; for any initial label box, obtaining the coordinate information corresponding to the initial label box includes: if the initial label box is the horizontal label box, determining the width, height, center point horizontal coordinate and center point vertical coordinate of the horizontal label box as the coordinate information of the horizontal label box; and if the initial label box is the rotated label box, determining the width, height, center point horizontal coordinate, center point vertical coordinate and rotation angle of the rotated label box as the coordinate information of the rotated label box.

[0010] In an embodiment, the label box inspection model is trained according to the following manner: obtaining each preference data sample in a preference data set, wherein the preference data sample includes a target training image, positive preference information and negative preference information; inputting the preference data sample into a multi-modal neural network model for iterative training, in any iterative training process, obtaining a current loss function value based on the positive preference information and the negative preference information; if the loss function value meets a preset requirement or the number of iterations meets a preset requirement, stopping iteration, taking the multi-modal neural network model that has been trained as the label box inspection model, and if the loss function value does not meet the preset requirement and the number of iterations does not meet the preset requirement, updating the model parameters of the multi-modal neural network model.

[0011] In an embodiment, the preference dataset is constructed according to the following manner: a plurality of target training image sets are obtained, a plurality of standard annotation boxes in each of the target training image sets are determined, a training annotation box set is generated based on coordinate information of the standard annotation boxes, wherein the standard annotation boxes are used to annotate target objects in any target training image; an area intersection-over-union ratio of any standard annotation box and each training annotation box in the corresponding training annotation box set is determined, and the training annotation box set is sorted in descending order according to the area intersection-over-union ratio; a preference data sample of the standard annotation box is constructed based on the target training image corresponding to the standard annotation box and the sorted training annotation box set, and a set of the preference data sample of each standard annotation box is taken as the preference dataset.

[0012] In an embodiment, generating the training annotation box set based on the coordinate information of the standard annotation box comprises: obtaining a plurality of coordinate information of the standard annotation box, performing random offset on each of the coordinate information, and generating a training annotation box based on the random offset result; and repeating the random offset on each of the coordinate information of the standard annotation box multiple times to construct the training annotation box set according to the generated plurality of training annotation boxes, wherein the training annotation boxes include the standard annotation box.

[0013] In an embodiment, the standard annotation box includes a horizontal standard annotation box and a rotated standard annotation box; obtaining a plurality of coordinate information of the standard annotation box and performing random offset on each of the coordinate information comprises: obtaining a plurality of first coordinate information of the horizontal standard annotation box, the first coordinate information including a width, a height, a center point horizontal coordinate and a center point vertical coordinate of the horizontal annotation box, randomly sampling in a preset variable set for any first coordinate information, and offsetting the first coordinate information based on the random sampling value; obtaining a plurality of second coordinate information of the rotated standard annotation box, the second coordinate information including a width, a height, a center point horizontal coordinate, a center point vertical coordinate and a rotation angle of the rotated annotation box, randomly sampling in a preset variable set for any second coordinate information, and offsetting the second coordinate information based on the random sampling value.

[0014] In an embodiment, constructing the preference data sample of the standard annotation frame based on the target training image corresponding to the standard annotation frame and the training annotation set after sorting includes: for any two training annotation frames in the training annotation set, determining a positive preference annotation frame and a negative preference annotation frame in the two training annotation frames according to the sorting order of the two training annotation frames in the training annotation set; determining the coordinate information of the positive preference annotation frame as positive preference information and the coordinate information of the negative preference annotation frame as negative preference information, and taking the target training image, the positive preference information, and the negative preference information as one of the preference data samples.

[0015] The second aspect of the present application provides a remote sensing image annotation frame intelligent generation device for on-orbit computing, the device comprising: an initial annotation unit configured to obtain a target remote sensing image set to be annotated, and perform preliminary annotation on the target remote sensing image set to obtain a plurality of initial annotation frames of the target remote sensing image set, wherein the initial annotation frame is used to annotate a target object in any target remote sensing image; a model analysis unit configured to obtain a target remote sensing image corresponding to any initial annotation frame and coordinate information, and input the target remote sensing image and the coordinate information into an annotation frame verification model, wherein the annotation frame verification model is obtained by pre-training based on a preference data set, and any preference data sample of the preference data set comprises at least positive preference information and negative preference information; and an optimized annotation unit configured to determine whether the initial annotation frame belongs to an alignment defect annotation frame by using the annotation frame verification model, and if it is determined that the initial annotation frame belongs to an alignment defect annotation frame, re-annotate the target remote sensing image corresponding to the initial annotation frame to obtain a target annotation frame of the target object.

[0016] The third aspect of the present application provides a computer device, comprising: a memory and a processor, which are communicatively connected with each other, and the memory stores computer instructions, and the processor executes the computer instructions to perform the remote sensing image annotation frame intelligent generation method for on-orbit computing according to the first aspect.

[0017] The technical scheme provided by one or more embodiments of the present application automatically evaluates the alignment quality of the labeling box through a labeling box inspection model, quickly screens out the labeling boxes with alignment defects for re-labeling, thereby improving the generation quality and inspection efficiency of the labeling box of the remote sensing image. Specifically, the coordinate information of the target remote sensing image and the initial labeling box is input into the labeling box inspection model, the labeling box inspection model performs multi-modal analysis on the image information and the coordinate text information, and outputs the alignment quality score of the labeling box to determine whether re-labeling is needed, thereby quickly identifying the labeling box with low alignment quality, improving the inspection efficiency of the labeling box of the remote sensing image, and keeping the labeling box of the target object at a high alignment degree, effectively improving the alignment quality of the labeling box of the remote sensing image target detection. In addition, the labeling box inspection model is pre-trained based on the preference data sample with positive preference information and negative preference information, so that the labeling box inspection model can learn the features of high-quality labeling boxes, thereby more accurately identifying the labeling box with alignment defects.

[0018] As can be seen, the technical scheme provided by the present application can automatically and quickly identify the labeling box with low alignment quality, improve the inspection efficiency and accuracy of the labeling box of the remote sensing image, and thereby improve the generation quality of the labeling box of the remote sensing image. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical scheme in the specific embodiments of the present application or the prior art, the drawings needed in the description of the specific embodiments or the prior art will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings based on these drawings without creative labor.

[0020] Figure 1 A method step diagram of a remote sensing image labeling box intelligent generation method for on-orbit computing provided by an embodiment of the present application;

[0021] Figure 2 A step schematic diagram of training a labeling box inspection model provided by an embodiment of the present application;

[0022] Figure 3 A step schematic diagram of constructing a preference data set provided by an embodiment of the present application;

[0023] Figure 4 A structure schematic diagram of a remote sensing image labeling box intelligent generation device for on-orbit computing provided by an embodiment of the present application;

[0024] Figure 5 A structure schematic diagram of a computer device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0025] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0026] In addition, the descriptions involving "first", "second" and the like in the present application are only for the purpose of description, and cannot be understood as indicating or implying the relative importance of the indicated technical features or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can be explicitly or implicitly included at least one of the features. In the description of the embodiments of the present application, unless otherwise specified, the meaning of "multiple" is two or more. In addition, the use of "based on" or "according to" means openness and inclusiveness, because the process, step, calculation or other action based on one or more described conditions or values can be based on additional conditions or values beyond the described values in practice.

[0027] With the continuous development of on-orbit computing technology, the demand for high-quality labeled data sets for training of remote sensing image target detection models for on-orbit computing is increasing. In the process of remote sensing image target detection for on-orbit computing, due to the characteristics of high resolution and large field of view of remote sensing images, and the problems of large scale difference, variable posture, complex background interference and the like of targets, the labeled boxes in the data set used by the remote sensing image target detection model for on-orbit computing in the training process need to strictly frame the actual physical boundary of the target, ensure the high alignment of the labeled box and the target area, and thus ensure the effectiveness of the subsequent model training for learning the features of the target.

[0028] However, in the actual remote sensing image data set labeling process, the current mainstream manual labeling method and automatic tool labeling method both have obvious limitations. Manual labeling relies on professional labeling personnel to manually draw labeled boxes for target objects, which has low labeling efficiency and human error. The automatic tool labeling is limited by the complexity and diversity of targets in remote sensing images, and the generated labeled boxes are often difficult to accurately match the actual area of the target. Therefore, it is necessary to correct through the labeled box alignment quality inspection and optimization link, but the traditional quality inspection method is still mainly based on manual sampling review. In the labeling scene of large-scale remote sensing data sets, manual sampling review is often difficult to cover all the data, and it is easy to miss the labeled samples with alignment problems. This inspection method is extremely low in efficiency, thereby leading to poor quality of labeled boxes.

[0029] Therefore, one or more embodiments of the present application provide a remote sensing image annotation box intelligent generation method and device for on-orbit computing, which can automatically and quickly identify low-alignment-quality annotation boxes, and improve the inspection efficiency, accuracy and generation quality of remote sensing image annotation boxes. Furthermore, sampling high-quality annotated remote sensing image datasets to train a remote sensing image target detection model for on-orbit computing can improve the performance of the remote sensing image target detection model.

[0030] Referring to Figure 1 One embodiment of the present application provides a remote sensing image annotation box intelligent generation method for on-orbit computing, which can include the following steps:

[0031] S1: Obtain a target remote sensing image set to be annotated, and perform preliminary annotation on the target remote sensing image set to obtain a plurality of initial annotation boxes, wherein the initial annotation boxes are used to annotate target objects in any target remote sensing image.

[0032] S3: Obtain a target remote sensing image and coordinate information corresponding to any initial annotation box, and input the target remote sensing image and the coordinate information into an annotation box inspection model, wherein the annotation box inspection model is obtained by pre-training based on a preference dataset, and any preference data sample of the preference dataset includes at least positive preference information and negative preference information.

[0033] S5: Determine whether the initial annotation box belongs to an alignment defect annotation box by using the annotation box inspection model, and if it is determined that the initial annotation box belongs to an alignment defect annotation box, re-annotate the target object corresponding to the initial annotation box, and use the re-annotated annotation box as the target annotation box of the target object.

[0034] The target remote sensing image set is a remote sensing image that needs to be detected and annotated, and contains one or more target objects to be annotated. The target objects in the target remote sensing image are preliminarily annotated to generate an initial annotation box for each target object. The annotation box can be understood as a rectangular bounding box used to frame the target object, which contains information such as width, height, and center point coordinates. The preliminary annotation process can use manual annotation or automatic tool annotation to annotate the target objects in each target remote sensing image in the target remote sensing image set to obtain a plurality of initial annotation boxes. The initial annotation box may not be completely aligned with the actual boundary of the target object, and needs to be further evaluated and optimized.

[0035] The target remote sensing image corresponding to the above annotation box is a target remote sensing image where the annotation is located, and the position information of the annotation box in the target remote sensing image is the coordinate information corresponding to the above annotation box. The coordinate information can include the length, height, center point coordinate, center point rotation angle, etc. of the annotation box. The above preference dataset is a dataset used to train the annotation box verification model, which contains a plurality of preference data samples, each sample including at least positive preference information and negative preference information. The positive preference information represents a relatively high alignment quality annotation box sample, and the negative preference information represents a relatively low alignment quality annotation box sample.

[0036] In this embodiment, before using the annotation box verification model, it needs to be trained using the preference dataset to learn how to distinguish annotation boxes of different alignment qualities. Specifically, the annotation box verification model can be trained based on a neural network model, and the preference dataset is the key to training the annotation box verification model. The positive preference information and the negative preference information in the preference dataset provide rich and distinct learning samples for the model. The annotation box verification model can automatically determine the alignment quality of the annotation box by learning the positive preference information and the negative preference information, thereby improving the recognition accuracy of the annotation box verification model. The annotation box verification model can quickly and accurately identify low alignment quality annotation boxes in actual application, and then the annotation personnel can optimize or re-annotate the low alignment quality annotation boxes, thereby improving the generation quality and verification efficiency of remote sensing image annotation boxes.

[0037] For example, a neural network model supporting multi-modal input is selected as the annotation box verification model to be trained, so that the annotation box verification model can process image data and text data at the same time. Specifically, a convolutional neural network is used to extract image features of the target remote sensing image. The image features can reflect the visual features of the target object in the image, such as shape, texture, color, etc. A text embedding technology (such as word embedding, position embedding, etc.) is used to convert the coordinate information of the annotation box into text features. The text features can reflect the geometric position information of the annotation box. The image features and the text features are fused to form a comprehensive feature vector. The fusion method can be simple splicing, or a more complex attention mechanism or fusion network.

[0038] The above alignment defect annotation box can be understood as an annotation box with poor alignment quality, i.e., the annotation box has a significant deviation from the actual boundary of the target object. The annotation box verification model can determine whether the initial annotation box belongs to the alignment defect annotation box by comparing the output confidence score with a preset threshold. For example, the model generates an alignment quality score, and compares the alignment quality score of the annotation box with a preset threshold. If the alignment quality score is lower than the threshold, it is determined that the annotation box is an alignment defect annotation box.

[0039] In the embodiment, the annotation box inspection model is used to identify the low alignment quality annotation box that needs to be relabeled, and ensures that all target object annotation boxes are high-quality target annotation boxes. If it is determined that a certain initial annotation box belongs to an alignment defect annotation box, the labeling personnel needs to relabel the target object corresponding to the annotation box, and replace the original initial annotation box with the newly generated target annotation box. It should be noted that the above target annotation box needs to be a high-quality alignment annotation box, otherwise it needs to be relabeled again to generate a target annotation box. The above target annotation box is well aligned with the actual boundary of the target object, can more accurately frame the target object, thereby improving the generation quality of the annotation box, reducing the necessity of manual inspection, and improving the labeling efficiency.

[0040] Based on the above idea, the technical scheme provided by the embodiment automatically evaluates the alignment quality of the annotation box through the annotation box inspection model, quickly screens out the alignment defect annotation box for relabeling, thereby improving the generation quality and inspection efficiency of the remote sensing image annotation box. The annotation box inspection model fuses image information and coordinate text information for multi-modal analysis to determine whether the annotation box labeled in the initial stage needs to be relabeled, thereby quickly identifying the low alignment quality annotation box for relabeling, improving the inspection efficiency of the remote sensing image annotation box, and keeping the annotation box of the target object at a high alignment degree, thereby effectively improving the alignment quality of the remote sensing image target detection annotation box. In addition, the annotation box inspection model is pre-trained based on the preference data sample with positive preference information and negative preference information, so that the annotation box inspection model can learn the features of high-quality annotation boxes, automatically and quickly identify low alignment quality annotation boxes, improve the inspection efficiency and accuracy of remote sensing image annotation boxes, and thereby improve the generation quality of remote sensing image annotation boxes.

[0041] In one embodiment, the alignment quality score generated by the annotation box inspection model is used to determine whether the annotation box is an alignment defect annotation box. Specifically, for any initial annotation box, the annotation box inspection model generates an alignment quality score for the initial annotation box, and the alignment quality score reflects the degree of alignment between the annotation box and the actual boundary of the target object. Further, it is determined whether the alignment quality score is less than a preset alignment threshold. If the alignment quality score is less than the preset alignment threshold, it is determined that the initial annotation box belongs to an alignment defect annotation box, and the target object corresponding to the annotation box needs to be relabeled. If the alignment quality score is greater than or equal to the preset alignment threshold, it is determined that the initial annotation box does not belong to an alignment defect annotation box, and the initial annotation box is directly used as the target annotation box of the corresponding target object.

[0042] In an embodiment, the alignment quality score generated by the annotation box verification model is used to determine whether the annotation box is an alignment defect annotation box, which realizes the automation and objectivity of alignment quality evaluation, can quickly identify annotation boxes with low alignment quality, so that the annotator can directly re-annotate these annotation boxes with low alignment quality without checking all annotation boxes one by one. At the same time, the alignment quality of different annotation boxes can be compared through a unified threshold, avoiding the subjectivity of manual judgment. The alignment quality score is automatically generated by the annotation box verification model, which realizes the automatic evaluation of the alignment quality of the annotation box, reduces the necessity of manual inspection, and thus improves the generation quality and inspection efficiency of the remote sensing image annotation box, and further improves the performance of the remote sensing image target detection model for on-orbit computing.

[0043] The technical solution provided by the embodiment realizes the automation and objectivity of alignment quality evaluation, can quickly identify annotation boxes with low alignment quality, so that the annotator can directly re-annotate these annotation boxes with low alignment quality without checking all annotation boxes one by one. At the same time, the alignment quality of different annotation boxes can be compared through a unified threshold, avoiding the subjectivity of manual judgment. The alignment quality score is automatically generated by the annotation box verification model, which realizes the automatic evaluation of the alignment quality of the annotation box, reduces the necessity of manual inspection, and thus improves the generation quality and inspection efficiency of the remote sensing image annotation box, and further improves the performance of the remote sensing image target detection model for on-orbit computing.

[0044] In an embodiment, the initial annotation box includes a horizontal annotation box and a rotated annotation box. For any initial annotation box, the corresponding coordinate information is obtained according to the type of the initial annotation box. Specifically, if the initial annotation box is a horizontal annotation box, the width, height, center point horizontal coordinate and center point vertical coordinate of the horizontal annotation box are determined as the coordinate information of the horizontal annotation box, and if the initial annotation box is a rotated annotation box, the width, height, center point horizontal coordinate, center point vertical coordinate and rotation angle of the rotated annotation box are determined as the coordinate information of the rotated annotation box.

[0045] The technical solution provided by the embodiment ensures that the model can process different types of annotation boxes and obtain detailed coordinate information by obtaining the corresponding coordinate information according to the type of the initial annotation box. By obtaining the detailed coordinate information of different annotation boxes, the model can adapt to target objects of different shapes and directions, thereby more accurately evaluating the alignment quality of the annotation box, helping to identify and optimize alignment defect annotation boxes with low alignment quality, and enhancing the flexibility and adaptability of the model in different application scenarios.

[0046] In an embodiment, please refer to Figure 2 The annotation box verification model is trained based on a pre-constructed preference data set, so that it can accurately identify alignment defect annotation boxes with poor alignment quality. Specifically, the training process is performed according to the following steps S51-S55:

[0047] S51: Obtain each preference data sample in the preference dataset, wherein the preference data sample includes the target training image, positive preference information and negative preference information.

[0048] S53: Input the preference data sample into the multimodal neural network model for iterative training. During any iterative training process, obtain the current loss function value based on the positive preference information and the negative preference information.

[0049] S55: If the loss function value meets the preset requirements or the number of iterations meets the preset conditions, then stop the iteration and use the currently trained multimodal neural network model as the bounding box verification model. If the loss function value does not meet the preset requirements and the number of iterations does not meet the preset conditions, then update the model parameters of the multimodal neural network model.

[0050] The aforementioned preference dataset includes multiple preference data samples, each containing a target training image, positive preference information, and negative preference information. The positive preference information is the coordinate information of a bounding box sample with relatively high alignment quality, while the negative preference information is the coordinate information of a bounding box sample with relatively low alignment quality. This multimodal neural network model can simultaneously process image data and text data, i.e., the target training image and the coordinate information of the training bounding boxes.

[0051] During any iteration of training, the coordinates of the target training image and the bounding boxes are input into the multimodal neural network model. Image features and text features are extracted separately, and then fused to form a comprehensive feature vector, which is mapped to the network layer to output an alignment quality score. In each iteration, the model calculates a loss function value based on positive and negative preference information. Commonly used loss functions include pairwise loss functions and contrastive loss functions. If the loss function value meets a preset requirement (e.g., converges to a stable value), or the total number of iterations meets a preset number of iterations, the iteration training stops. If the loss function value does not meet the preset requirement and the total number of iterations does not meet the preset number of iterations, the model parameters are updated, and training continues.

[0052] For example, for positive preference information in a preference data sample and negative preference information The loss function value can be calculated in the following way: ,in, For positive preference information The corresponding alignment quality score, negative preference information The corresponding alignment quality score. The goal of the loss function is to make the alignment quality score of positive samples higher than that of negative samples.

[0053] The technical scheme provided by the embodiment can realize accurate identification of low-quality alignment defect annotation boxes through constructing an annotation box verification model trained based on a preference data set. The multi-modal neural network model iteratively learns the target training image and annotation box coordinate information of each preference data sample, and in each iteration round, optimizes the model parameters according to the loss function values generated by the positive preference information and the negative preference information, so that the finally trained annotation box verification model can accurately identify the alignment defect standard boxes that need to be re-labeled, thereby improving the generation quality and verification efficiency of the remote sensing image annotation boxes.

[0054] In one embodiment, please refer to Figure 3 Before training the annotation box verification model, a preference data set for training is constructed, and the preference data set is constructed based on the following steps:

[0055] S61: Obtain a plurality of target training image sets, and determine a plurality of standard annotation boxes in each of the target training image sets, and generate a training annotation box set based on the coordinate information of the standard annotation boxes, wherein the standard annotation boxes are used to annotate target objects in any target training image.

[0056] S63: Determine the area intersection ratio of any standard annotation box and each training annotation box in the corresponding training annotation box set, and sort the training annotation box set in descending order according to the area intersection ratio.

[0057] S65: Construct the preference data sample of the standard annotation box based on the target training image corresponding to the standard annotation box and the sorted training annotation set, and take the set of preference data samples of each standard annotation box as the preference data set.

[0058] The target training image set includes a plurality of groups of remote sensing images containing target objects, and the standard annotation box is a standard box with high alignment quality pre-selected to accurately frame the target object in the remote sensing image. There is a training annotation box set corresponding to any standard annotation box, and the training standard box set is a set of annotation boxes generated according to the coordinate information of the standard annotation box, which is used to simulate different alignment quality annotation boxes of the target object of the standard annotation box. For example, for each standard annotation box, the coordinate information such as width, height, and center point coordinate is randomly offset to generate a plurality of new annotation boxes.

[0059] In the embodiment, the standard training frame set is sorted according to the area intersection ratio of each standard annotation frame and each training annotation frame. Specifically, for each standard annotation frame, the area intersection ratio of the standard annotation frame and each annotation frame in the training annotation frame set is calculated, and the training annotation frame set is sorted in descending order of the area intersection ratio, so that the annotation frame with a higher area intersection ratio is arranged in front, and the annotation frame with a lower area intersection ratio is arranged in back. The area intersection ratio can reflect the alignment quality of the training annotation frame, so that for any two training annotation frames, the training annotation frame with higher alignment quality and the training annotation frame with lower alignment quality can be determined according to the sorting order.

[0060] In the embodiment, for any standard annotation frame, a preference data sample of the standard annotation frame is constructed according to any two training annotation frames in the training annotation frame set. Specifically, for each standard annotation frame, any two annotation frames are selected from the sorted training annotation frame set, the annotation frame with a higher area intersection ratio is taken as a positive preference annotation frame, and the annotation frame with a lower area intersection ratio is taken as a negative preference annotation frame to construct a preference data sample, which contains a target training image, a positive preference annotation frame and a negative preference annotation frame. The set of all preference data samples is taken as a preference data set. By generating the training annotation frame set, calculating the area intersection ratio and sorting, and constructing the preference data sample, the preference data set can provide rich learning samples for the annotation frame verification model and help the model learn to distinguish annotation frames with different alignment qualities.

[0061] In the embodiment, on the basis of the above step S61, the coordinate information of the standard annotation frame is randomly offset to generate the training annotation frame set based on the coordinate information. Specifically, the following steps are performed:

[0062] S611: Obtain a plurality of coordinate information of the standard annotation frame, randomly offset each coordinate information, and generate a training annotation frame based on the random offset result;

[0063] S612: Repeat the random offset of each coordinate information of the standard annotation frame multiple times to construct the training annotation frame set according to the generated plurality of training annotation frames, wherein the training annotation frame set includes the standard annotation frame.

[0064] In the embodiment, new annotation boxes are generated by randomly changing the coordinate information of the standard annotation boxes to simulate annotation boxes with different alignment qualities. Specifically, the width, height, center point horizontal coordinate, center point vertical coordinate, and other information of the standard annotation box are extracted, and each coordinate information is randomly offset. For example, for any standard annotation box, the width and height are randomly increased or decreased within a certain range, the horizontal coordinate, vertical coordinate, and rotation angle of the center point are randomly offset within a certain range, and a new annotation box is generated according to the randomly offset coordinate information. For any standard annotation box, the above random offset process is repeated multiple times to generate multiple training annotation boxes, and all generated training annotation boxes are combined into a set to form a training annotation box set.

[0065] It should be noted that the standard annotation box is included in the training annotation box set to enable the model to learn the characteristics of high-quality annotation boxes. By randomly offsetting the coordinate information of the standard annotation box to generate annotation boxes with different alignment qualities, the model can learn the characteristics of high-quality and low-quality annotation boxes, thereby more accurately evaluating the alignment quality of the annotation boxes in actual applications, thereby improving the generation quality and testing efficiency of the remote sensing image annotation boxes.

[0066] In the embodiment, on the basis of the above step S611, since the standard annotation box includes horizontal standard annotation boxes and rotated standard annotation boxes, multiple coordinate information of the standard annotation box is obtained, and random offset is performed on each coordinate information, including the following two cases:

[0067] Case one: multiple first coordinate information of the horizontal standard annotation box is obtained, the first coordinate information includes the width, height, center point horizontal coordinate, and center point vertical coordinate of the horizontal annotation box, random sampling is performed on the preset variable set for any first coordinate information, and the first coordinate information is offset based on the random sampling value.

[0068] Case two: multiple second coordinate information of the rotated standard annotation box is obtained, the second coordinate information includes the width, height, center point horizontal coordinate, center point vertical coordinate, and rotation angle of the rotated annotation box, random sampling is performed on the preset variable set for any second coordinate information, and the second coordinate information is offset based on the random sampling value.

[0069] Specifically, the preset variable set is a parameter set containing multiple random offset values, which is used to control the degree of offset. For any standard annotation box, a value is randomly sampled from the preset variable set for each coordinate information, and the coordinate information is offset according to the randomly sampled value. After all the coordinate information is offset, new coordinate information is formed, and the standard box represented by the new coordinate information is taken as a training standard box. For example, for a training standard box of a standard annotation box, the width of the training standard box is the original width of the standard annotation box plus a randomly sampled value, and the length of the training standard box is the original length of the standard annotation box plus a randomly sampled value.

[0070] In the embodiment, on the basis of the step S65, the preference data sample of the standard annotation box is constructed based on the target training image corresponding to the standard annotation box and the sorted training annotation set, and specifically includes: for any two training annotation boxes in the training annotation set, determining a positive preference annotation box and a negative preference annotation box in the two training annotation boxes according to the sorting order of the two training annotation boxes in the training annotation set, determining the coordinate information of the positive preference annotation box as positive preference information, and determining the coordinate information of the negative preference annotation box as negative preference information, and taking the target training image, the positive preference information and the negative preference information as a preference data sample.

[0071] In any two training annotation boxes of the training annotation set, the positive preference information represents the position information of the training annotation box with relatively high alignment quality, i.e. the position information of the training annotation box with high sorting order, and the negative preference information represents the position information of the training annotation box with relatively low alignment quality, i.e. the position information of the training annotation box with low sorting order.

[0072] In the embodiment, for any standard annotation box, the preference data sample of the standard annotation box is constructed according to the sorting order of any two training annotation boxes. Specifically, for each standard annotation box, any two annotation boxes are selected in the sorted training annotation box set, the annotation box with high sorting order is taken as the positive preference annotation box, and the annotation box with low sorting order is taken as the negative preference annotation box to construct a preference data sample containing the target training image, the positive preference annotation box and the negative preference annotation box. The above process is repeated multiple times to take the set of all constructed preference data samples as a preference data set, so that the preference data set can provide rich learning samples for the annotation box verification model and help the model to learn to distinguish annotation boxes with different alignment qualities.

[0073] The technical scheme provided by the embodiment trains the annotation box verification model by constructing a preference data set, thereby significantly improving the generation quality and verification efficiency of the annotation box of the remote sensing image. Specifically, first, a set of training annotation boxes containing different alignment qualities is generated by randomly offsetting the coordinate information of the high-alignment-quality standard annotation box selected in advance. Then, the area intersection ratio of the standard annotation box and each training annotation box is calculated and sorted in descending order, and positive preference information and negative preference information are generated according to the order relationship of any two training annotation boxes. The preference data samples are combined with the corresponding target training images to form a complete preference data set. By simulating various alignment deviations in real scenarios through random offsetting, and by considering both horizontal and rotated annotation box types, the model can effectively learn to distinguish the features of annotation boxes with different alignment qualities and different types, thereby providing a data basis for the subsequent annotation box verification model to accurately identify the alignment defect annotation boxes, and improving the annotation quality and verification efficiency of the remote sensing image.

[0074] The application also provides an embodiment of constructing a preference data set, which is performed according to the following steps:

[0075] S101: For any standard annotation box of a target object of a certain type in a target training image set, a training annotation box corresponding to the standard annotation box is obtained by randomly offsetting each coordinate parameter of the standard annotation box.

[0076] If the standard annotation box is a horizontal annotation box, let the width and height of the standard annotation box be (w, h), and the coordinate of the center point of the standard annotation box in the image be (i, j). Sample 1 time from each of the preset variable set , and assume that four sample values are obtained. Then, a training annotation box corresponding to the standard annotation box is constructed, and the width and height of the training annotation box are , the coordinate of the center point of the training annotation box in the image is .

[0077] If the standard annotation box is a rotated annotation box, let the width and height of the standard annotation box be (w, h), the coordinate of the standard annotation box in the image be (i, j), and the rotation angle be . Sample 1 time from each of the preset variable set , and assume that five sample values are obtained. Then, a training annotation box corresponding to the standard annotation box is constructed, and the width and height of the training annotation box are , the coordinate of the center point of the training annotation box in the image is , and the rotation angle is .

[0078] It should be noted that any random variable value in the above-mentioned preset variable set is a random variable value .), can be subject to any value range containing the probability distribution of positive and negative numbers, or can be directly equal to a constant.

[0079] S102: Repeat step S101, so that each standard annotation box corresponds to a plurality of training annotation boxes to form a corresponding training annotation box set , m represents the number of training annotation boxes, and the area intersection ratio of each training annotation box and the standard annotation box is calculated to obtain the area intersection ratio of the annotation box ( ) and the standard annotation box , and the training annotation box set is sorted in descending order according to the area intersection ratio, so that the sorted For any , there is .

[0080] S103: Any standard annotation box of the target object in the target training image set is combined with its corresponding sorted training annotation box set and the image corresponding to the standard annotation box to form a data sample , wherein is the standard annotation box, and then any two training annotation boxes are selected in the above data sample, and the positional relationship of the two training annotation boxes is constructed as a preference data sample of the current standard annotation box according to the order relationship and the image , the above process is repeated to obtain multiple preference data samples of the current standard annotation box , and then the preference data samples of multiple standard annotation boxes are obtained to form a preference data set.

[0081] The application also provides an embodiment of a training process of an annotation box verification model, which is performed according to the following steps:

[0082] S201: Construct a standard box detection model that supports simultaneous input of text data and image data and outputs a scalar score, and the input and output process of the above standard box detection model can be represented as , wherein represents the score output by the model, i.e. the alignment quality score, represents the image data input into the model, i.e. the target training image, represents the text data input into the model, i.e. the preference information.

[0083] S202: For each preference data sample in the preference data set , for any , a preference learning sample is constructed, wherein, is positive bias information, is text composed of coordinate information of a training bounding box is negative bias information, is text composed of coordinate information of a training bounding box .

[0084] S203: Based on the preference learning sample constructed in step S202, iteratively train the standard bounding box detection model until the model converges. The standard bounding box detection model can be constructed from scratch using a multi-modal neural network, or the output layer of a trained multi-modal visual language model can be modified to obtain it.

[0085] Referring to Figure 4 , the application also provides a remote sensing image bounding box intelligent generation device for on-orbit computing, which comprises:

[0086] An initial labeling unit 100 is configured to obtain a target remote sensing image set to be labeled, and perform preliminary labeling on the target remote sensing image set to obtain a plurality of initial bounding boxes of the target remote sensing image set, wherein the initial bounding box is used to label a target object in any target remote sensing image.

[0087] A model analysis unit 200 is configured to obtain a target remote sensing image corresponding to any initial bounding box and coordinate information, and input the target remote sensing image and the coordinate information into a bounding box verification model, wherein the bounding box verification model is obtained by pre-training based on a preference data set, and any preference data sample of the preference data set comprises at least positive bias information and negative bias information.

[0088] An optimized labeling unit 300 is configured to determine whether the initial bounding box belongs to an alignment defect bounding box by using the bounding box verification model, and if it is determined that the initial bounding box belongs to an alignment defect bounding box, re-label the target remote sensing image corresponding to the initial bounding box to obtain a target bounding box of the target object.

[0089] Wherein,

[0090] In one embodiment, the initial labeling unit 100 is specifically configured to preliminarily label the target object in the target remote sensing image, and generate an initial bounding box for each target object. The initial bounding box can be a horizontal bounding box or a rotated bounding box. The preliminary labeling process can be performed by manual labeling or automatic tool labeling.

[0091] ​In an embodiment, the model analysis unit 200 is specifically configured to use the annotation box inspection model to perform multi-modal analysis and evaluation on the annotation box based on the target remote sensing image and the coordinate information of the annotation box, generate image features based on the target remote sensing image, generate text features based on the coordinate information, fuse the image features and the text features to obtain a comprehensive feature vector, input the fused comprehensive feature vector into a scoring network inside the model, and output an alignment quality score by the scoring network. The alignment quality of the current annotation box is evaluated according to the alignment quality score.

[0092] In an embodiment, the optimization annotation unit 300 is specifically configured to, for any initial annotation box, generate an alignment quality score of the initial annotation box by using the annotation box inspection model, determine whether the alignment quality score is less than a preset alignment threshold, if the alignment quality score is less than the preset alignment threshold, determine that the initial annotation box belongs to an alignment defect annotation box, re-annotate the target object corresponding to the annotation box, and if the alignment quality score is greater than or equal to the preset alignment threshold, determine that the initial annotation box does not belong to the alignment defect annotation box, and directly use the initial annotation box as the target annotation box of the corresponding target object.

[0093] Further function descriptions of the above-mentioned modules and units are the same as those of the corresponding embodiments, and will not be repeated here.

[0094] The on-orbit computing-oriented remote sensing image annotation box intelligent generation device in the embodiment of the present application is presented in the form of a functional unit. The unit here refers to an ASIC (Application Specific Integrated Circuit, Application Specific Integrated Circuit) circuit, a processor and a memory executing one or more software or fixed programs, or other devices that can provide the above functions.

[0095] Please refer to Figure 5 , Figure 5 is a structural schematic diagram of a computer device provided in the embodiment of the present application, as Figure 5As shown, the computer device includes one or more processors 10, memory 20, and interfaces 30 for external devices such as a keyboard and a mouse and peripheral devices such as disk devices or other storage devices. One or more busses 10 can be used to implement the interface between the various internal and external components and can be implemented using any one or more of a variety of bus technologies including a System bus, PCI, SCSI, AGP, Super- I / O bus, etc. Furthermore, various buses can be used in front side buses, back side buses, and other bus configurations based on any bus or messaging technology known to those skilled in the art. Figure 5 The processor 10 is used in the embodiments below as an example.

[0096] The processor 10 can be a central processing unit, a network processor, or a combination thereof. The processor 10 can further include a hardware chip. The hardware chip can be an application specific integrated circuit, a programmable logic device, or a combination thereof. The programmable logic device can be a complex programmable logic device, a field programmable logic device, a general array logic, or any combination thereof.

[0097] The memory 20 stores instructions that can be executed by the at least one processor 10, so that the at least one processor 10 implements the method shown in the above embodiments.

[0098] The memory 20 can include a program region and a data region. The program region can store an operating system and an application program required by at least one function. The data region can store data created according to the use of the computer device, etc. In addition, the memory 20 can include a high-speed random access memory, and can further include a non-transitory memory such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state memory device. In some alternative embodiments, the memory 20 can optionally include a memory disposed remotely with respect to the processor 10, and these remote memories can be connected to the computer device through a network. Examples of the network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0099] The memory 20 can include a volatile memory such as a random access memory, and can also include a non-volatile memory such as a flash memory, a hard disk, or a solid state disk. The memory 20 can further include a combination of the above-mentioned kinds of memories.

[0100] The computer device further includes a communication interface 30 for communication with other devices or communication networks.

[0101] The apparatus or units illustrated in the above embodiments can be specifically implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, the computer can be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0102] For the ease of description, the above apparatus is described in functions as various units respectively. Of course, functions of the units can be implemented in a same or multiple software and / or hardware when implementing the present application.

[0103] Those skilled in the art should understand that the embodiments of the present application can be provided as a method or an apparatus. Therefore, the present application can adopt a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt a form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.

[0104] The present application is described with reference to flowcharts and / or block diagrams of the methods, apparatus, and computer devices according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one flow or multiple flows and / or blocks Figure 1 The apparatus that implements the functions specified in one block or multiple blocks.

[0105] These computer program instructions can also be stored in a computer readable storage medium that can direct the computer or other programmable data processing devices to work in a specific manner, so that the instructions stored in the computer readable storage medium produce a product including an instruction apparatus that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one flow or multiple flows and / or blocks Figure 1 The apparatus that implements the functions specified in one block or multiple blocks.

[0106] These computer program instructions can also be loaded to the computer or other programmable data processing devices, so that a series of operation steps are performed on the computer or other programmable data processing devices to produce a computer implemented process, so that the instructions executed on the computer or other programmable data processing devices provide a process for implementing the functions specified in the flowcharts and / or block diagrams.Figure 1 one or more processes and / or functions specified in one or more blocks. Figure 1 one or more blocks or any combination thereof.

[0107] It should also be noted that the terms "comprising," "comprises," "including," "includes" or other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises a... " does not, without more constraints, exclude the existence of additional identical elements in the process, method, article, or apparatus that comprises the recited element.

[0108] Each of the embodiments in the present specification is described in a progressive manner, and the same or similar parts between the embodiments can be mutually referred to, and each of the embodiments mainly explains the difference from other embodiments. In particular, 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.

[0109] The above only describes the embodiments of the present application and is not intended to limit the present application. The present application can have various modifications and changes for those skilled in the art. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the scope of claims of the present application.

[0110] Although the embodiments of the present application are described in conjunction with the drawings, various modifications and changes can be made by those skilled in the art without departing from the spirit and scope of the present application, and such modifications and changes fall within the scope defined by the appended claims.

Claims

1. A method for intelligent generation of remote sensing image labeling boxes for on-orbit computing, characterized in that, The method comprises: acquiring a target remote sensing image set to be labeled, and performing preliminary labeling on the target remote sensing image set to obtain a plurality of initial labeling boxes, wherein the initial labeling box is used for labeling a target object in any target remote sensing image; acquiring a target remote sensing image and coordinate information corresponding to any initial labeling box, and inputting the target remote sensing image and the coordinate information into a labeling box inspection model, wherein the labeling box inspection model is obtained by pre-training based on a preference data set, and any preference data sample in the preference data set comprises at least positive preference information and negative preference information; determining whether the initial labeling box belongs to an alignment defect labeling box by using an alignment quality score generated by the labeling box inspection model, and if it is determined that the initial labeling box belongs to an alignment defect labeling box, re-labeling the target object corresponding to the initial labeling box, and taking the labeling box obtained by re-labeling as a target labeling box of the target object, wherein the alignment defect labeling box is a labeling box deviating from the actual boundary of the target object.

2. The method of claim 1, wherein, Determining whether the initial labeling box belongs to an alignment defect labeling box by using the labeling box inspection model comprises: generating an alignment quality score of the initial labeling box by using the labeling box inspection model, and determining whether the alignment quality score is less than a preset alignment threshold; if the alignment quality score is less than the preset alignment threshold, it is determined that the initial labeling box belongs to an alignment defect labeling box, and if the alignment quality score is greater than or equal to the preset alignment threshold, it is determined that the initial labeling box does not belong to the alignment defect labeling box.

3. The method according to claim 1 or 2, characterized in that, The method further comprises: if it is determined that the initial labeling box does not belong to the alignment defect labeling box, taking the initial labeling box as a target labeling box of the corresponding target object.

4. The method of claim 1, wherein, The types of the initial labeling box include a horizontal labeling box and a rotated labeling box, and for any initial labeling box, acquiring coordinate information corresponding to the initial labeling box comprises: if the initial labeling box is the horizontal labeling box, determining the width, height, center point horizontal coordinate and center point vertical coordinate of the horizontal labeling box as the coordinate information of the horizontal labeling box; if the initial labeling box is the rotated labeling box, determining the width, height, center point horizontal coordinate, center point vertical coordinate and rotation angle of the rotated labeling box as the coordinate information of the rotated labeling box.

5. The method of claim 1, wherein, The labeling box inspection model is trained according to the following manner: acquiring each preference data sample in a preference data set, wherein the preference data sample comprises a target training image, positive preference information and negative preference information; inputting the preference data sample into a multi-modal neural network model for iterative training, and in any iterative training process, obtaining a current loss function value based on the positive preference information and the negative preference information; if the loss function value meets a preset requirement, stopping iteration, taking the current trained multi-modal neural network model as the labeling box inspection model, and if the loss function value does not meet the preset requirement, updating the model parameters of the multi-modal neural network model.

6. The method of claim 1, wherein, The preference data set is constructed according to the following manner: obtaining a plurality of target training image sets, and determining a plurality of standard annotation boxes in each of the target training image sets, generating a training annotation box set based on coordinate information of the standard annotation boxes, wherein the standard annotation boxes are used to annotate target objects in any target training image; determining an area intersection-over-union ratio of any standard annotation box and each training annotation box in the corresponding training annotation box set, and performing descending order sorting on the training annotation box set according to the area intersection-over-union ratio; constructing a preference data sample of the standard annotation box based on a target training image corresponding to the standard annotation box and the sorted training annotation box set, and taking a set of preference data samples of each standard annotation box as the preference data set.

7. The method of claim 6, wherein, The generating of the training annotation box set based on the coordinate information of the standard annotation boxes comprises: obtaining a plurality of coordinate information of the standard annotation boxes, randomly offsetting each of the coordinate information, and generating a training annotation box based on the random offset result; repeating the random offsetting of each of the coordinate information of the standard annotation boxes multiple times to construct the training annotation box set according to the generated plurality of training annotation boxes, wherein the training annotation boxes include the standard annotation boxes.

8. The method of claim 7, wherein, The types of the standard annotation boxes include horizontal standard annotation boxes and rotated standard annotation boxes; obtaining a plurality of coordinate information of the standard annotation boxes, randomly offsetting each of the coordinate information, and generating a training annotation box based on the random offset result; obtaining a plurality of first coordinate information of the horizontal standard annotation boxes, the first coordinate information including width, height, center point horizontal coordinate and center point vertical coordinate of the horizontal standard annotation box, randomly sampling in a preset variable set for any first coordinate information, and offsetting the first coordinate information based on the random sampling value; obtaining a plurality of second coordinate information of the rotated standard annotation boxes, the second coordinate information including width, height, center point horizontal coordinate, center point vertical coordinate and rotation angle of the rotated standard annotation box, randomly sampling in a preset variable set for any second coordinate information, and offsetting the second coordinate information based on the random sampling value.

9. The method of claim 6, wherein, The constructing of the preference data sample of the standard annotation box based on the target training image corresponding to the standard annotation box and the sorted training annotation box set comprises: for any two training annotation boxes in the training annotation box set, determining a positive preference annotation box and a negative preference annotation box in the two training annotation boxes according to the sorting order of the two training annotation boxes in the training annotation box set; determining coordinate information of the positive preference annotation box as positive preference information, and determining coordinate information of the negative preference annotation box as negative preference information, and taking the target training image, the positive preference information and the negative preference information as one preference data sample.

10. An on-orbit computing-oriented remote sensing image label box intelligent generation device, characterized in that, The device comprises: an initial annotation unit configured to obtain a target remote sensing image set to be annotated, and perform preliminary annotation on the target remote sensing image set to obtain a plurality of initial annotation boxes of the target remote sensing image set, wherein the initial annotation boxes are used to annotate target objects in any target remote sensing image; A model analysis unit is configured to obtain a target remote sensing image corresponding to any of the initial annotation boxes and coordinate information, and input the target remote sensing image and the coordinate information into an annotation box verification model. The annotation box verification model is pre-trained based on a preference data set. Any preference data sample of the preference data set includes at least positive preference information and negative preference information. An optimization annotation unit is configured to determine, based on an alignment quality score generated by the annotation box verification model, whether the initial annotation box belongs to an alignment defect annotation box. If it is determined that the initial annotation box belongs to the alignment defect annotation box, the target remote sensing image corresponding to the initial annotation box is re-annotated to obtain a target annotation box of the target object. The alignment defect annotation box is an annotation box deviating from an actual boundary of the target object.

11. A computer device, comprising: Comprise: A memory and a processor are communicatively connected to each other. The memory stores computer instructions. The processor executes the computer instructions to perform the method of any one of claims 1 to 9.

Images