Video redirection method and system based on video saliency sorting

Through the significant object sorting model based on Mask R-CNN and graph neural network, combined with LOESS smoothing technology, the problem that existing video redirection methods cannot accurately capture the significant differences in objects is solved, and the significance area is realized during video redirection The accuracy of ensuring retention and scene continuity improves the user experience.

CN119048948AActive Publication Date: 2024-11-29SOUTHEAST UNIV

Patent Information

Application Number
CN202410935102.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-12
Publication Date
2024-11-29
Estimated Expiration
2044-07-12

AI Technical Summary

Technical Problem

Existing video redirection methods cannot accurately capture the significant differences between different objects in the foreground, resulting in biases on background or objects with less significant significance during the redirection process, affecting the overall quality of the video content and the user's viewing experience.

Method used

Using a significant object sorting model based on Mask R-CNN and graph neural network, the instance features are extracted through attention mechanism and position protection, the significance value of each instance is compared and sorted, the weight is allocated according to the degree of significance, and combined with the instance The position weighting is used to obtain the cropping center, and the video frame is cropped using the LOESS method to smoothly crop the time series of the cropping center to realize video redirection.

Benefits of technology

Retain the significant area to the greatest extent, avoid artifacts and timing inconsistencies, ensure the continuity of the scene and the video quality, and eliminate phenomena such as artifacts and distortions that affect the visual experience.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119048948A_ABST
    Figure CN119048948A_ABST
Patent Text Reader

Abstract

The invention discloses a video redirection method and system based on video saliency ranking. The method comprises the following steps: firstly, acquiring a video saliency object ranking data set at least comprising each instance bounding box, a mask result and saliency ranking; the method comprises the following steps: constructing a significant object sorting model based on a Mask R-CNN and a graph neural network, distinguishing significant objects and non-significant objects in an image through the Mask R-CNN, extracting instance features through an attention mechanism and position protection attention (PPA), inputting the features into the graph neural network, comparing significance values of all instances, and sorting; and calculating a saliency center as a cutting center, obtaining the size of a cutting frame according to a target length-width ratio, finally smoothing the time sequence of the cutting center by using an LOESS method, and cutting the video frame according to the smoothed cutting center and the size of the cutting frame to obtain a redirected video. According to the method, the salient region can be reserved to the maximum extent, and the continuity of the scene is ensured.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of computer vision technology, and in particular to a video salient object sorting task and a video redirection technology, and mainly to a video redirection method and system based on video saliency sorting. Background Art

[0002] With the development of multimedia technology, people can watch videos on different devices such as smartphones and tablets. Video retargeting adjusts videos to different aspect ratios to fit different terminals. The purpose of image redirection is to preserve visually salient areas for a better visual experience. Based on this, image retargeting has achieved great success in the past through methods such as seam carving, cropping, and warping. However, simply applying these methods to video retargeting will lead to severe temporal inconsistencies, greatly affecting the human visual experience.

[0003] The smartVi dCrop method uses saliency detection for cropping to achieve video redirection. However, this method only distinguishes between foreground and background, ignoring the saliency differences between different instances in the foreground. Therefore, this method often leads to biases towards background or less saliency objects during the redirection process. Specifically, in the foreground scene, different objects may have different importance and attention, but smartVi dCrop fails to accurately capture these differences, which makes it unable to accurately retain truly important salient objects when processing complex scenes. As a result, some less important backgrounds or less saliency foreground objects may be mistakenly considered to be the focus of redirection, affecting the overall quality of the video content and the user's viewing experience.

[0004] Recently, the booming salient object ranking task (SOR) provides an important method to solve the saliency difference. The salient object ranking task aims to rank objects according to their saliency in an image. Saliency usually refers to the degree to which an object attracts attention in a visual scene. The main goal of this task is to identify and prioritize those objects that are most visually salient in order to better understand the structure and focus of the scene. In the image field, the relative saliency of instances is predicted by capturing multi-scale spatial contrasts, mainly including inter-instance contrast, instance-to-instance local contrast, and instance-to-global contrast. In the video field, the temporal characteristics of instances should also be taken into account, including motion, etc. This puts forward new requirements for saliency ranking. However, the current video salient object ranking method still cannot accurately capture the motion information of the instance and cannot achieve accurate saliency ranking. Therefore, how to better utilize the results of saliency ranking to achieve video redirection has become a direction that technicians in this field are competing to study and solve. Summary of the invention

[0005] The present invention is aimed at the problem that the existing redirection process cannot guarantee the quality of the visual salient area, and there are problems such as distortion, artifacts, inconsistent timing, and errors in extracting salient areas. A video redirection method and system based on video saliency sorting is provided. First, a bounding box containing at least each instance, a mask result and a saliency ranking are obtained to obtain a video salient object sorting dataset; a salient object sorting model based on Mask R-CNN and a graph neural network is constructed, and salient objects and non-salient objects in the image are distinguished by Mask R-CNN. After the features of the instance are extracted by the attention mechanism and position protection attention (PPA), the features are input into the graph neural network, and the saliency values ​​of each instance are compared and sorted to obtain the saliency degree of each instance; a weight is assigned to each instance according to its saliency degree, and a cropping center is obtained by weighting according to the weight and the instance position; the size of the cropping box is obtained according to the target aspect ratio, and finally the time series of the cropping center is smoothed by the LOESS method, and the video frame is cropped according to the size of the smoothed cropping center and the cropping box to obtain the redirected video. The method of the present invention can retain the salient area to the greatest extent and ensure the continuity of the scene.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is: a video redirection method based on video significance sorting, which specifically includes the following steps:

[0007] S1, obtaining a data set: obtaining a video salient object ranking data set, wherein the data set at least includes a bounding box, a mask result and a saliency ranking of each instance;

[0008] S2, constructing a salient object ranking model: The model is constructed based on Mask R-CNN and graph neural network. Mask R-CNN is used to distinguish salient objects from non-salient objects in the image. After extracting the features of the instance through the attention mechanism and position-protected attention (PPA), the features are input into the graph neural network, and the saliency value of each instance is compared and ranked to obtain the saliency degree of each instance.

[0009] S3, calculate the saliency center: assign a weight to each instance according to its saliency, and obtain the cropping center based on the weight and instance position;

[0010] S4, determining the cropping frame: obtaining the size of the cropping frame according to the target aspect ratio, specifically: making the long side occupy the entire original image and scaling the short side to the target ratio;

[0011] S5, smoothing: Use the LOESS method to smooth the time series of the cropping center, and crop the video frame according to the size of the smoothed cropping center and the cropping box to obtain the redirected video.

[0012] As an improvement of the present invention, the graph neural network in step S2 is used to mine the spatiotemporal relationship of instances;

[0013] At the spatial level, for local feature extraction, the bounding box area of ​​each instance is proportionally enlarged by 1 times to obtain the bounding box of the local area of ​​the instance, and the features within the bounding box are extracted using ROI Align.

[0014] For global feature extraction, the entire video frame is divided into 3×3 parts, and the features of each part are extracted through the average pooling layer;

[0015] At the temporal level, for the extraction of trajectory features, the bounding box area of ​​the current frame instance is proportionally enlarged by 1 times and the features in the bounding box are extracted using ROI Align at the corresponding position of the adjacent frame as the trajectory features of the instance.

[0016] As another improvement of the present invention, the loss function of the salient object sorting model in step S2 includes the object detection loss L 1 and the saliency ranking loss L 2 , where the object detection loss L 1 is the classification loss L cls , regression box loss L box and the mask loss L mask The sum of the significant ranking loss L 2 Specifically:

[0017]

[0018] Among them, {q 1 ,q 2} is any pair of instances; is the significance score; β q is the weight coefficient; is the combination number formula, which represents the number of all instance pairs, where N is the number of instances.

[0019] As another improvement of the present invention, the weight coefficient β q Specifically:

[0020]

[0021] in, For the instance {q 1 ,q 2}’s ranking prediction, Represents an instance {o 1 ,o 2}’s real ranking results, is the combination number formula, which represents the number of all instance pairs, where N is the number of instances.

[0022] As another improvement of the present invention, in step S3, the calculation method of the clipping center is specifically as follows:

[0023]

[0024] Among them, (x i ,y i ) is the center coordinate of each instance, r i is the significance ranking, r i =1,2…,r i The larger the value, the stronger the significance.

[0025] In order to achieve the above object, the present invention also adopts a technical solution: a video redirection system based on video significance sorting includes a computer program, and when the computer program is executed by a processor, the steps of any of the above methods are implemented.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] (1) The present invention realizes video redirection based on salient object sorting. Compared with the previous methods based on seam carving and distortion, the present invention eliminates the occurrence of artifacts by cropping. Compared with other cropping-based methods, the present invention uses salient object sorting instead of saliency detection to prevent the influence of the less salient foreground and background on the cropped area, and can retain the visual salient area to the maximum extent.

[0028] (2) The video saliency ranking model framework proposed in this paper first segments instances based on Mask R-CNN, then enhances instance features through the attention mechanism, and uses the position-preserving attention module to introduce position and scale as saliency priors; subsequently, a spatiotemporal graph neural network is constructed to jointly optimize multi-scale spatial and temporal contrast cues to generate instance-level saliency features; finally, a fully connected neural network is used to infer saliency scores, and combined with the results of instance segmentation, the final saliency ranking result is obtained, which is more accurate.

[0029] (3) The graph neural network of this method is used to mine the spatiotemporal relationship of instances. The saliency of instances in the spatiotemporal relationship is considered at the spatial level and the temporal level. In the spatial level, the interactive relationship between instances, the contrast relationship between instances and local instances, and the contrast relationship between instances and the global instance are considered; in the temporal level, the mutual relationship between instances at different times and the motion characteristics of each instance are considered, and the spatiotemporal relationship is jointly optimized to capture the spatiotemporal saliency. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1It is a flow chart of the steps of the video redirection method based on video significance sorting of the present invention;

[0031] Figure 2 It is a structural schematic diagram of the video redirection framework of the present invention;

[0032] Figure 3 It is a structural schematic diagram of the video saliency ranking model of the present invention;

[0033] Figure 4 is a schematic diagram of the result effect of video significance sorting in Example 1 of the present invention;

[0034] Figure 5 It is a comparison chart of the effects of different redirection methods in the test example of the present invention. DETAILED DESCRIPTION

[0035] The present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0036] Example 1

[0037] Video redirection methods based on video salient object sorting, such as Figure 1 and Figure 2 As shown, a video salient object sorting data set and a video salient detection data set are used to train a video salient object sorting model. Through the trained video salient object sorting model, the salient instance information of each frame can be obtained, including the bounding box, mask and saliency ranking of each instance. Then, a weight is assigned to each instance based on its saliency ranking. The stronger the saliency, the greater the weight. Based on this weight, the global saliency center is obtained as the cropping center and a cropping box is generated based on the target aspect ratio. Finally, the position of the obtained cropping center is smoothed in time series to make the entire scene coherent. Specifically, the following steps are included:

[0038] Step S1: Obtain a video salient object ranking dataset for training a video salient model. According to the requirements of the video salient object ranking task, for each frame in the video, the information that needs to be annotated includes: bounding boxes, masks, and saliency levels of all salient instances.

[0039] The video salient object ranking dataset used is extracted from the public datasets RVSOD and DAVSOD. In this embodiment, the data in RVSOD and DAVSOD are processed as follows:

[0040] 1) For RVSOD, the bounding box and mask information of each instance are missing. The instance-level mask is deducted from the existing image-level mask and the bounding box is generated. For the saliency level of the instance, a saliency level is assigned to each instance based on the density of fixations provided by RVSOD. The division of the training set and the test set is the same as that of RVSOD.

[0041] 2) For DAVSOD, we obtain instance-level annotations in the same way as for RVSOD. In addition, some videos in the DAVSOD dataset contain only a single salient instance and are therefore not suitable for saliency ranking. Therefore, we choose to discard this part of the data and divide the remaining videos into two categories: some frames contain multiple salient instances and all frames contain multiple salient instances. Finally, each type of video is divided into a training set and a test set in a ratio of 4:1.

[0042] Step S2: Use Mask R-CNN and graph neural network to build a salient object sorting model.

[0043] The present invention uses Mask R-CNN to segment instances and mines the salient features of instances through graph neural networks. The video salient object sorting model includes an object detection module (Mask R-CNN) and a salient sorting module based on a graph convolutional neural network. The two modules are trained in series to ensure that the entire model can be trained and predicted end-to-end. In the detection module, the model only distinguishes between salient objects and non-salient objects in the image without identifying them. In the sorting module, different instances are sorted by obtaining salient values ​​based on the features of the detected instances combined with the spatiotemporal relationship. The entire salient sorting model is as follows: Figure 3 As shown in the figure: After extracting instance features through Mask R-CNN, the features of the instance are fully extracted through the attention mechanism and position-protected attention (PPA), and then these features are input into the graph neural network. Finally, the saliency value of each instance is compared and ranked. The detailed steps include:

[0044] S21: Use Mask R-CNN for instance segmentation, in which only salient objects and non-salient objects are distinguished;

[0045] S22: In instance segmentation, Mask R-CNN uses ROI Align to extract Region Proposal features. Considering the impact of shallow features such as position information and size on saliency, the extracted instance features are positionally encoded and the attention mechanism is used to extract instance saliency information.

[0046] S23: In order to fully extract the saliency information of the instance, a graph convolutional neural network is constructed to mine the spatiotemporal relationship of the instance: At the spatial level, the main considerations are the interaction between instances, the interaction between instances and their local features, and the interaction between instances and the global scene. For local feature extraction, the bounding box area of ​​each instance is proportionally enlarged by 1 times to obtain the bounding box of the local area of ​​the instance, and then ROI Align is used to extract the features in the bounding box. For feature extraction, the entire frame is divided into 3×3 parts, and each feature is extracted through an average pooling layer. At the temporal level, the main considerations are the interaction between different instances and the impact of the instance's trajectory information on the instance. For the extraction of trajectory features, we proportionally enlarge the bounding box area of ​​the instance in the current frame by 1 times and use ROI Align at the corresponding position of the adjacent frame to extract the features in the bounding box as the instance's trajectory features;

[0047] S24: The loss function of constructing a video salient object ranking model based on graph convolutional neural network mainly consists of two parts: the first is the loss of object detection, including classification loss, regression box loss, and mask loss. The specific calculation formula is as follows:

[0048] L 1 =L cls +L box +L mask

[0049] For saliency ranking, consider all instance pairs. For any pair of instances {q 1 ,q 2}, where q 1 is less significant than q 2 The predicted significance scores are The goal is to reduce and increase the value Therefore, the specific calculation formula is as follows:

[0050]

[0051] where β q is a weight coefficient, the purpose of which is to assign larger weights to pairs with larger ranking differences and smaller weights to pairs with similar rankings, so as to explicitly optimize instances with extreme ranks. The specific calculation formula is as follows:

[0052]

[0053] In this embodiment, the parameter settings of the video saliency ranking model based on graph neural network are: the learning rate is 5e-6, Adam is used as the optimizer, the batch size is set to 1, the maximum number of iterations is set to 200,000, and the learning rate decays to 1 / 10 of the current learning rate at the 80,000th and 150,000th iterations.

[0054] Figure 4 The effect of the present invention on the task of sorting salient objects in a video is shown in the figure. The first two rows are the input video frames and salient instances, and the third, fourth, fifth and sixth rows are the results of the method with only spatial salient signals, the method with only global temporal signals and spatial signals, the method with only instance-level temporal signals and spatial signals, and the method of the present invention. Figure 4 It can be seen from the above that before the introduction of the method for extracting time series features in the present invention, the model cannot accurately mine the time series significance of the instance. The present invention introduces the trajectory features of the instance to enable the model to accurately extract the significant features of the instance, achieving the best results so far.

[0055] Step S3, obtain the cropping center according to the predicted saliency result: The predicted saliency result includes the position of each saliency instance and its saliency degree. A weight is assigned to each instance according to its saliency degree, and the cropping center is obtained by weighting the weight and the instance position. The center coordinates of each instance are (x i ,y i ) The significance ranking is r i (where r i =1,2…,and r The larger the i, the stronger the significance). Then the cropping center is:

[0056]

[0057] Step S4, obtaining the size of the cropping frame according to the target aspect ratio, for the target length and width f w / f h , we make the long side fill the original image and scale the short side to the target ratio. For example, if the target length and width f w / f h =4 / 5, we keep the image height unchanged and only crop it on the x-axis.

[0058] Step S5, using the LOESS method to smooth the time series of the cropping center, and cropping the video frame according to the smoothed cropping center and the size of the cropping frame to obtain the redirected video.

[0059] In summary, the method of the present invention uses the public data sets RVSOD and DAVSOD to train a graph neural network-based model to sort video salient objects. Based on the extracted salient instances and their salient strengths, a suitable cropping center is selected and a suitable cropping area is selected in combination with the required aspect ratio. Finally, a smoothing method is used to make the final result coherent. The method of the present invention overcomes the problem of poor visual experience in the visual salient area during the redirection process. Through the salient object sorting module, all salient instances can be obtained and non-salient foregrounds or backgrounds can be isolated. Furthermore, through sorting, a weight is assigned to the salient instances, so that in the case of extreme redirection aspect ratios, the present invention can retain the most salient visual areas. Not only that, through cropping, the occurrence of artifacts, distortions, and other phenomena that affect human visual experience is eliminated.

[0060] Test Case

[0061] Performance test of video salient object sorting based on graph neural network: To verify the effectiveness of the saliency sorting module in the present invention, this test example performs a performance test on the video salient object sorting task and compares it with multiple advanced models.

[0062] The evaluation indicators of the video salient object sorting task are as follows:

[0063]

[0064] Where S(x,y) represents the predicted pixel grayscale value, and GT(x,y) represents the actual pixel grayscale value. The MAE indicator is used to measure the absolute error between the predicted image and the label image. The lower the value, the better the model.

[0065] SSA-SOR=Pearson-Correlation(Ranks pre ,Ranks gt )

[0066] Pearson-Correlation means calculating the Pierce correlation coefficient, Ranks pre ,Ranks gt Represent the predicted and actual saliency level sequences respectively. SA-SOR is the most advanced metric for ranking salient objects. Compared with the previous SOR and SSOR, SA-SOR can reflect both segmentation and ranking performance.

[0067] The comparison results of the video salient object ranking method based on graph neural network in the present invention and the quantitative comparison method of other two image salient object rankings and another advanced model of video salient object ranking are shown in the following table:

[0068]

[0069] It can be seen from the above table that the video salient object ranking method based on graph neural network in the present invention surpasses the most advanced methods in both data sets and two indicators, showing extremely strong competitiveness.

[0070] Performance test of the video redirection method based on video saliency sorting: In order to verify the effectiveness of the video redirection method of the present invention, this test example performs a performance test on the video redirection task and compares it with multiple advanced models.

[0071] Since there is no universal theorem evaluation metric for video redirection, visualization results are chosen for qualitative comparison.

[0072] The comparison results of the video redirection method based on salient object sorting in the present invention and the other two video redirection methods are shown in the figure. Figure 5 As shown: the first column is the input video frame, and the next six columns are divided into two parts: the results of aspect ratios of 2:3 and 4:5 are shown, where the first, second and third columns are the seam carving method, the saliency-based cropping method and the method of the present invention. Figure 5 It can be seen that the seam carving-based method often causes instance artifacts and affects the visual experience (yellow box in the figure); the saliency-based cropping method is easily interfered by weakly salient objects and background information and generates erroneous cropping frames. On the contrary, the method of the present invention uses instance-level saliency levels for redirection to achieve the best visual effect.

[0073] First, the present invention adopts a video salient object sorting method to extract the salient areas of the video frame, takes advantage of the graph neural network to model feature interactions, and uses spatial comparison graphs and temporal comparison graphs to fully explore the spatiotemporal salient clues of the instance, thus achieving the best video salient object sorting result at present. Secondly, the present invention adopts a video redirection method based on video salient object sorting. The method uses cropping to eliminate artifacts and a series of results that affect human visual experience. Moreover, compared with the previous method based on video saliency, the method based on video salient object sorting fully isolates the background and foreground and eliminates the influence of weakly salient foreground, making the cropping result more accurate. The present invention explains and verifies the rationality of the proposed video redirection method based on video saliency sorting through a large number of comparative experiments.

[0074] The video redirection method based on video salient object sorting provided by the present invention can convert the input video into the target aspect ratio and retain the salient area to improve the visual experience. Specifically in the industrial world, with the development of multimedia technology and the popularization of a series of devices with different aspect ratios such as computers, mobile phones, watches, and tablets, the advantages of the redirection method can be maximized to obtain a model that meets the needs of the enterprise field and is close to supervised pre-training.

[0075] It should be noted that the above content only illustrates the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications all fall within the protection scope of the claims of the present invention.

Claims

1. A video redirection method based on video saliency ranking, characterized in that: The specific steps include: S1, obtaining a data set: obtaining a video salient object ranking data set, wherein the data set at least includes a bounding box, a mask result and a saliency ranking of each instance; S2, building a salient object ranking model: The model is built based on Mask R-CNN and graph neural network. Mask R-CNN is used to distinguish salient objects from non-salient objects in the image. After extracting the features of the instance through the attention mechanism and position-protected attention (PPA), the features are input into the graph neural network, and the saliency value of each instance is compared and ranked to obtain the saliency degree of each instance. S3, calculate the saliency center: assign a weight to each instance according to its saliency, and obtain the cropping center based on the weight and instance position; S4, determining the cropping frame: obtaining the size of the cropping frame according to the target aspect ratio, specifically: making the long side occupy the entire original image and scaling the short side to the target ratio; S5, smoothing: Use the LOESS method to smooth the time series of the cropping center, and crop the video frame according to the size of the smoothed cropping center and the cropping box to obtain the redirected video.

2. The video redirection method based on video saliency ranking as claimed in claim 1, characterized in that: The graph neural network in step S2 is used to mine the spatiotemporal relationship of instances; At the spatial level, for local feature extraction, the bounding box area of ​​each instance is proportionally enlarged by 1 times to obtain the bounding box of the local area of ​​the instance, and ROI Align is used to extract the features within the bounding box; For global feature extraction, the entire video frame is divided into 3×3 parts, and the features of each part are extracted through the average pooling layer; At the temporal level, for the extraction of trajectory features, the bounding box area of ​​the current frame instance is proportionally enlarged by 1 times and ROI Align is used at the corresponding position of the adjacent frame to extract the features in the bounding box as the trajectory features of the instance.

3. The video redirection method based on video saliency ranking as claimed in claim 2, characterized in that: The loss function of the salient object ranking model in step S2 includes object detection loss L1 and salient ranking loss L2, where object detection loss L1 is the classification loss L2. cls , regression box loss L box and the mask loss L mask The significance ranking loss L2 is specifically: Among them, {q1,q2} is any pair of instances; is the significance score; β q is the weight coefficient; is the combination number formula, which represents the number of all instance pairs, where N is the number of instances.

4. The video redirection method based on video significance ranking as claimed in claim 3, characterized in that: The weight coefficient β q Specifically: in, represents the ranking prediction for instance {q1,q2}, represents the actual ranking result of instance {o1,o2}, is the combination number formula, which represents the number of all instance pairs, where N is the number of instances.

5. The video redirection method based on video saliency ranking as claimed in claim 1, characterized in that: In step S3, the calculation method of the cutting center is specifically as follows: Among them, (x i ,y i ) is the center coordinate of each instance, r i is the significance ranking, r i =1,2…,r i The larger the value, the stronger the significance.

6. A video redirection system based on video saliency ranking, comprising a computer program, characterized in that: When the computer program is executed by a processor, the steps of any of the above methods are implemented.

Citation Information

Patent Citations

  • Method for reframing images of a video sequence, and apparatus for reframing images of a video sequence

    CN103533370A

  • Video behavior identification method based on mobile uncut network

    CN110188654A

  • Method capable of carrying out end-to-end training and simultaneously carrying out target detection and saliency sorting

    CN113420828A

  • Video saliency target detection method based on quality evaluation

    CN113936235A

Cited By

  • Subject person image segmentation method, apparatus and device, and medium

    WO2026179250A1