Rotor flow field determination method, device, equipment and storage medium

Abstract

The application provides a rotor flow field determination method, device and equipment and a storage medium, relates to the technical field of image data processing, and the method adopts a capsule layer to extract features of a rotor flow field image by using a pre-trained optical flow field prediction model for each frame of the rotor flow field image, obtains an attitude matrix, determines structural embedding features based on the attitude matrix, subtracts the structural embedding features of the first frame of the rotor flow field image from the structural embedding features of the second frame of the rotor flow field image, and obtains motion embedding information; and determines a rotor optical flow field based on the motion embedding information. The capsule layer can output features in the form of vectors, features are easier to extract, the optical flow field is easier to determine, and the motion information of an object, i.e. the motion embedding information, is obtained by calculating the difference between the structural embedding features of two frames of images, thereby avoiding the complexity of traditional correlation methods and reducing the complexity of the method while improving the accuracy of the optical flow field determination.

Description

Technical Field

[0001] This application relates to the field of image data processing technology, specifically to a method, apparatus, device, and storage medium for determining rotor flow field. Background Technology

[0002] When helicopters fly near the ground, the powerful downwash generated by the rotor can lift smoke or other particulate matter from the ground into the air, creating particle aggregation that obstructs visibility. This particle aggregation affects the pilot's visibility and the aerodynamic performance of the helicopter rotor, posing a serious threat to helicopter safety. These particles exhibit certain characteristics that change with the helicopter's flight attitude, and the particle characteristics also differ between different helicopter models. In relevant particle flow field experiments, the resulting particle aggregation is an unavoidable phenomenon that occurs under unsteady conditions and endangers helicopter operation. Its flow characteristics change continuously over time, and non-contact, continuously measurable measurement methods should be employed based on the actual situation.

[0003] Most methods for determining optical flow fields in related technologies are the traditional Local Optical Flow (LK) algorithm and Global Optical Flow (HS) algorithm. However, the LK algorithm is limited by the spatial consistency assumption and can only solve the velocity field of a finite number of pixels in the neighborhood, which may result in significant sparsity. Although the HS algorithm can generate a dense velocity field, its global smoothing assumption is significantly affected by changes in illumination and noise interference. Both algorithms estimate by minimizing energy, which is insufficient for solving large displacement scenes. These problems can lead to target detection errors, motion blur, or trajectory breaks during the solution process.

[0004] Significant progress has been made in optical flow estimation using deep learning, enabling optical flow calculations even in scenes with large displacements. Fischer et al. introduced FlowNetC, which uses a set of learned features to calculate correlations and refines the features at multiple scales. Additionally, the SpyNet network proposes overlaying two images to learn spatiotemporal features at multiple scales. PWC-Net uses the correlations between learned features to obtain disparity estimates at multiple scales and refines the resulting flow field hierarchically. However, these methods typically rely on correlated features or learning the relationships between a set of pixels in a given neighborhood for matching calculations. This over-reliance on hyperparameter settings can lead to the inability to find accurate matches, and the computational complexity is also high. Summary of the Invention

[0005] This application provides a method, apparatus, device, and storage medium for determining the rotor flow field, which can perform optical flow calculations on large displacement scenarios during particle flow field experiments, and can reduce the complexity of the method while improving the accuracy of optical flow field determination.

[0006] This application provides a method for determining the flow field of a rotor, including:

[0007] Acquire two adjacent rotor flow field images; the two adjacent rotor flow field images are the first rotor flow field image and the second rotor flow field image, respectively, and the second rotor flow field image is the image following the first rotor flow field image.

[0008] Based on two adjacent frames of rotor flow field images and a pre-trained optical flow field prediction model, the rotor optical flow field is obtained. The pre-trained optical flow field prediction model is used to extract features from each frame of rotor flow field image using a capsule layer to obtain an attitude matrix. Based on the attitude matrix, structural embedding features are determined. The attitude matrix represents the geometric state of particulate matter features, and the structural embedding features represent the attitude information of particulate matter and other objects. The structural embedding features of the second frame of rotor flow field image are subtracted from the structural embedding features of the first frame of rotor flow field image to obtain motion embedding information. The rotor optical flow field is determined based on the motion embedding information.

[0009] Optionally, the step of extracting features from the rotor flow field image using capsule layers to obtain an attitude matrix includes:

[0010] Basic features are extracted from the rotor flow field image to obtain the basic features;

[0011] The first capsule layer is used to extract the first feature from the basic features to obtain the first capsule feature;

[0012] The second capsule layer is used to extract the second feature from the features of the first capsule to obtain the second capsule feature.

[0013] The features of the second capsule are classified to obtain the pose matrix.

[0014] Optionally, the basic feature extraction of the rotor flow field image to obtain basic features includes:

[0015] The rotor flow field image is input into a convolutional layer for basic feature extraction to obtain the basic features.

[0016] Optionally, determining the rotor optical flow field based on the motion embedding information includes:

[0017] The motion embedding information and the attitude matrix of the first frame rotor flow field image are spliced ​​together to obtain the splicing features;

[0018] The rotor optical flow field is determined based on the splicing characteristics.

[0019] Optionally, determining the rotor optical flow field based on the splicing features includes:

[0020] The spliced ​​features are upsampled using transposed convolution to obtain the rotor optical flow field.

[0021] Optionally, before obtaining the rotor optical flow field based on two adjacent frames of rotor flow field images and a pre-trained optical flow field prediction model, the method further includes:

[0022] The optical flow field prediction model is trained with the objective of minimizing the objective loss function, resulting in a pre-trained optical flow field prediction model. The objective loss function includes:

[0023] ;

[0024] ;

[0025] ;

[0026] in, Let be the target loss function. As the first weight, To activate loss, As the second weight, For the reconstruction loss, i represents other object categories in the two frames of rotor flow field images, and t represents the target object category in the two frames of rotor flow field images, where the target category is floating particles. The activation probability of the target category. represents the activation probability of other categories, and m is the value that controls the difference between activation probabilities. This is the first frame of the rotor flow field image. The image is obtained by upsampling and reconstructing the structure embedding features corresponding to the first frame of the rotor flow field image.

[0027] To achieve the above and other related objectives, this application provides a rotor flow field determination device, comprising:

[0028] The data acquisition module is used to acquire two adjacent frames of rotor flow field images; the two adjacent frames of rotor flow field images are the first frame rotor flow field image and the second frame rotor flow field image, respectively, and the second frame rotor flow field image is the image following the first frame rotor flow field image.

[0029] The data processing module is used to obtain the rotor optical flow field based on two adjacent frames of rotor flow field images and a pre-trained optical flow field prediction model. The pre-trained optical flow field prediction model is used to extract features from each frame of rotor flow field image using a capsule layer to obtain an attitude matrix. Based on the attitude matrix, structural embedding features are determined. The attitude matrix represents the geometric state of particulate matter features, and the structural embedding features represent the attitude information of particulate matter and other objects. The structural embedding features of the second frame of rotor flow field image are subtracted from the structural embedding features of the first frame of rotor flow field image to obtain motion embedding information. The rotor optical flow field is determined based on the motion embedding information.

[0030] Optionally, the data processing module also includes:

[0031] The first processing unit is used to extract basic features from the rotor flow field image to obtain basic features;

[0032] The second processing unit is used to perform first feature extraction on the basic features using the first capsule layer to obtain the first capsule features;

[0033] The third processing unit is used to extract second features from the first capsule features using the second capsule layer to obtain the second capsule features;

[0034] The fourth processing unit is used to classify the features of the second capsule to obtain the pose matrix.

[0035] To achieve the above and other related objectives, this application also provides an electronic device, the electronic device comprising:

[0036] One or more processors;

[0037] Memory used to store the executable program code of the processor;

[0038] The processor is configured to execute the program code to implement the above-described rotor flow field determination method.

[0039] To achieve the above and other related objectives, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a computer's processor, causes the computer to perform one or more of the aforementioned rotor flow field determination methods.

[0040] As described above, the rotor flow field determination method, apparatus, device, and storage medium provided in this application have the following beneficial effects:

[0041] This application discloses a method for determining the rotor flow field. This method uses a pre-trained optical flow field prediction model to extract features from each frame of the rotor flow field image using a capsule layer, obtaining an attitude matrix. Based on the attitude matrix, structural embedding features are determined. The structural embedding features of the first frame of the rotor flow field image are subtracted from the structural embedding features of the second frame to obtain motion embedding information. The rotor optical flow field is then determined based on this motion embedding information. The capsule layer can output features in vector form, making it easier to extract features to determine the optical flow field. Furthermore, by calculating the difference in structural embedding features between two frames, the motion information of the object, i.e., the motion embedding information, is obtained, avoiding the complexity of traditional correlation methods. This method can reduce complexity while improving the accuracy of optical flow field determination.

[0042] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0043] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. In the drawings:

[0044] Figure 1 This is a flowchart illustrating a rotor flow field determination method according to an exemplary embodiment of this application;

[0045] Figure 2 This is a structural block diagram of a rotor flow field determination device illustrated in an exemplary embodiment of this application. Detailed Implementation

[0046] The embodiments of this application will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be understood that the preferred embodiments are only for illustrating this application and are not intended to limit the scope of protection of this application.

[0047] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. Therefore, the drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0048] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the present application. However, it will be apparent to those skilled in the art that embodiments of the present application may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the present application.

[0049] Please see Figure 1 , Figure 1 This is a flowchart illustrating a rotor flow field determination method according to an exemplary embodiment of this application. (Reference) Figure 1 It can be seen that the method for determining the rotor flow field can include:

[0050] Step S110: Obtain two adjacent frames of rotor flow field images.

[0051] Among them, two adjacent rotor flow field images are the first rotor flow field image and the second rotor flow field image, respectively, and the second rotor flow field image is the image following the first rotor flow field image.

[0052] In one embodiment of this application, in a particle flow field experiment, two adjacent frames of rotor flow field images can be acquired. A particle flow field experiment typically refers to an experiment that artificially creates a multi-particle scenario to test or study visual capabilities, protective equipment performance, or equipment reliability under particle aggregation phenomena. Relevant videos of the particle flow field experiment can be acquired using a camera device, and frames can be extracted from these videos as rotor flow field images.

[0053] Step S120: Based on two adjacent frames of rotor flow field images and a pre-trained optical flow field prediction model, the rotor optical flow field is obtained. The pre-trained optical flow field prediction model is used to extract features from each frame of rotor flow field image using a capsule layer to obtain an attitude matrix. Based on the attitude matrix, structural embedding features are determined. The attitude matrix represents the geometric state of particulate matter features, and the structural embedding features represent the attitude information of particulate matter and other objects. The structural embedding features of the second frame of rotor flow field image are subtracted from the structural embedding features of the first frame of rotor flow field image to obtain motion embedding information. The rotor optical flow field is determined based on the motion embedding information.

[0054] In one embodiment of this application, two adjacent frames of rotor flow field images can be input into a pre-trained optical flow field prediction model to obtain the rotor optical flow field. The pre-trained optical flow field prediction model can be used to extract features from each input frame of rotor flow field image using a capsule layer to obtain an attitude matrix. Based on the attitude matrix, structural embedding features are determined. Based on the structural embedding features of the second and first frames of rotor flow field images, motion embedding information is obtained. The rotor optical flow field is then determined based on the motion embedding information. The rotor optical flow field can include the motion direction and velocity of each pixel.

[0055] Among them, the attitude matrix represents the geometric state of particulate matter features, and the structural embedding feature represents the attitude information of particulate matter and other objects.

[0056] In one embodiment of this application, for each frame of the rotor flow field image, a capsule layer can be used to extract features from the rotor flow field image to obtain an attitude matrix, and structural embedding features can be determined based on the attitude matrix. The capsule layer includes multiple corresponding capsules, and the capsule layer can output vectors that better match the features needed in the rotor optical flow field determination process. Motion embedding information can be obtained based on the structural embedding features of the second frame rotor flow field image and the first frame rotor flow field image. The motion embedding information can characterize the motion features of each gravel in the second frame rotor flow field image compared to the first frame rotor flow field image. The rotor optical flow field can be obtained by transposed convolution based on the motion embedding information.

[0057] It should be noted that determining the rotor optical flow field can predict the critical region and time window of cockpit visibility obstruction, and can also compare the particle flow field characteristics under different blade configurations to quantify the design improvement effect. The rotor optical flow field can also be used to assist users in making other effect judgments.

[0058] Subtracting the structural embedding features of the first rotor flow field image from the structural embedding features of the second rotor flow field image allows us to characterize the motion features of each feature in the second rotor flow field image compared to the first rotor flow field image in a particle flow field experiment. By calculating the difference in the object's structural embedding features between the two images, we can obtain the object's motion information, i.e., motion embedding information, thus avoiding the complexity of traditional correlation methods.

[0059] In one possible implementation, the process of extracting features from the rotor flow field image using capsule layers to obtain the attitude matrix may include: extracting basic features from the rotor flow field image to obtain basic features; extracting first features from the basic features using a first capsule layer to obtain first capsule features; extracting second features from the first capsule features using a second capsule layer to obtain second capsule features; and classifying the second capsule features to obtain the attitude matrix. Classifying the second capsule features yields the attitude matrix and activation probabilities.

[0060] It should be noted that the rotor flow field image can include a first frame and a second frame. The corresponding attitude matrices can be extracted from the first and second frames of the rotor flow field image, respectively. The rotor flow field image can be input into four convolutional layers with a kernel size of 3 and a RULE activation function to extract basic features. Then, a first shallow feature extraction can be performed on the basic features using a convolution with a kernel size of 3. A first capsule layer can then be used to extract the first feature from the basic features extracted in the first shallow feature extraction, resulting in the first capsule feature output. A second shallow feature extraction can then be performed on the first capsule feature using a convolution with a kernel size of 7. A second capsule layer can then be used to extract the second feature from the first capsule feature extracted in the second shallow feature extraction, resulting in the second capsule feature output. A convolution with a kernel size of 1 can then be used to classify the second capsule feature, obtaining features of floating particles and other objects in the image. Each feature is presented as an attitude matrix and activation probability. The feature with the highest activation probability is selected as the category to be further calculated, i.e., the target category.

[0061] In one possible implementation, the process of extracting basic features from the rotor flow field image can include: inputting the rotor flow field image into a convolutional layer for basic feature extraction. This can be achieved using four sequentially connected convolutional layers, each with a ReLU activation function and a kernel size of 3, to extract the basic features from the rotor flow field image.

[0062] Optionally, the process of determining structural embedding features based on the pose matrix may include using a fully connected layer to convert the pose matrix into structural embedding features. Through a fully connected layer, the obtained pose matrix can be converted into structural embedding features containing pose information of particles and other objects.

[0063] In one possible implementation, the process of determining the rotor optical flow field based on motion embedding information may include: stitching the motion embedding information and the attitude matrix of the first frame of the rotor flow field image to obtain stitched features; and determining the rotor optical flow field based on the stitched features. Stitching the motion embedding information and the attitude matrix of the first frame of the rotor flow field image can provide more sufficient information for determining the rotor optical flow field. The attitude matrix of the first frame of the rotor flow field image provides the corresponding attitude, and the motion embedding information provides the differences in the motion states of various categories of objects between adjacent frames. The stitched features obtained can include richer learnable information, thereby improving the accuracy of rotor optical flow field determination.

[0064] In one possible implementation, the process of determining the rotor optical flow field based on the stitched features may include: upsampling the stitched features using transposed convolution to obtain the rotor optical flow field. Transposed convolution, also known as deconvolution or fractionally strided convolution, is an upsampling operation in deep learning. Transposed convolution can expand the size of the stitched features to obtain the rotor optical flow field. The kernel size of the transposed convolution can be 4.

[0065] Before obtaining the rotor optical flow field based on two adjacent frames of rotor flow field images and a pre-trained optical flow field prediction model, the method further includes:

[0066] The optical flow field prediction model is trained with the objective of minimizing the objective loss function, resulting in a pre-trained optical flow field prediction model. The objective loss function includes:

[0067] ;

[0068] ;

[0069] ;

[0070] in, Let be the target loss function. As the first weight, To activate loss, As the second weight, For the reconstruction loss, i represents other object categories in the two frames of rotor flow field images, and t represents the target object category in the two frames of rotor flow field images, where the target category is floating particles. The activation probability of the target category. represents the activation probability of other categories, and m is the value that controls the difference between activation probabilities. This is the first frame of the rotor flow field image. The image is obtained by upsampling and reconstructing the structure embedding features corresponding to the first frame rotor flow field image. This indicates that the average of the data is calculated.

[0071] Where m is the difference between the activation probabilities, which can be set to 0.95. It can be used to train activation vectors, ensuring that the network can accurately activate the target category. It can be used to ensure that the network learns the correct pose matrix and to train the network to recover the original image from the pose matrix. It can be set to 0.05. It can be set to 2.5.

[0072] A training sample set can be obtained, which can include multiple training sample pairs. Each training sample pair can include a training sample and a sample label. The training sample can be a rotor flow field image of two adjacent frames, and the sample label can be the corresponding rotor optical flow field. A multi-particulate environment can be simulated through a 3D rendering engine to generate an image sequence with accurate optical flow ground truth.

[0073] It should be noted that the purpose of the first capsule layer is to transform the CNN feature map into vectorized capsules for representation. Multiple capsules are generated through convolution, and each capsule contains a pose matrix and activation probability. Each capsule represents the local low-level feature information of the receptive field of a small patch region.

[0074] The purpose of the second capsule layer is to aggregate the features of multiple first-layer capsules using a CNN with a large receptive field, capturing global deep features. This still contains multiple capsules, each containing a pose matrix and activation probabilities.

[0075] Traditional CNN models tend to lose information about the spatial relationships between objects, while capsule layers can preserve the pose information (position, probability of existence, etc.) of objects in a vector-like manner.

[0076] In summary, the rotor flow field determination method provided in this application effectively solves the problems of traditional methods and deep learning measurement methods in related technologies. First, multi-layer capsules model the object using continuous attitude matrices, addressing the sensitivity of deep learning networks to hyperparameter settings and the limited receptive field in neighborhood correlation matching. Second, subtracting the object attitude matrices from the features of two image frames yields motion information, significantly reducing computational complexity and improving the estimation capability of large-displacement optical flow. Then, the attitude matrix of the first frame rotor flow field image is used in conjunction with motion embedding information for image reconstruction and optical flow estimation. Image reconstruction ensures the network learns the correct attitude matrix, improving the model's generalization ability. Finally, applying the rotor flow field determination method provided in this application to rotor start-up and hovering states in a multi-particulate environment effectively estimates the global particulate flow field velocity.

[0077] Figure 2 This is a block diagram illustrating a rotor flow field determination device according to an exemplary embodiment of this application. Figure 2 As shown, the exemplary rotor flow field determination device 200 includes:

[0078] The data acquisition module 210 is used to acquire two adjacent frames of rotor flow field images; the two adjacent frames of rotor flow field images are the first frame rotor flow field image and the second frame rotor flow field image, respectively, and the second frame rotor flow field image is the frame following the first frame rotor flow field image.

[0079] The data processing module 220 is used to obtain the rotor optical flow field based on two adjacent frames of rotor flow field images and a pre-trained optical flow field prediction model. The pre-trained optical flow field prediction model is used to extract features from each frame of rotor flow field image using a capsule layer to obtain an attitude matrix. Based on the attitude matrix, structural embedding features are determined. The attitude matrix represents the geometric state of particulate matter features, and the structural embedding features represent the attitude information of particulate matter and other objects. The structural embedding features of the second frame of rotor flow field image are subtracted from the structural embedding features of the first frame of rotor flow field image to obtain motion embedding information. The rotor optical flow field is determined based on the motion embedding information.

[0080] In one embodiment of this application, the data processing module further includes:

[0081] The first processing unit is used to extract basic features from the rotor flow field image to obtain basic features;

[0082] The second processing unit is used to perform first feature extraction on the basic features using the first capsule layer to obtain the first capsule features;

[0083] The third processing unit is used to extract second features from the first capsule features using the second capsule layer to obtain the second capsule features;

[0084] The fourth processing unit is used to classify the features of the second capsule to obtain the pose matrix.

[0085] In one embodiment of this application, the data processing module is further configured to:

[0086] The rotor flow field image is input into a convolutional layer for basic feature extraction to obtain the basic features.

[0087] In one embodiment of this application, the data processing module is further configured to:

[0088] The motion embedding information and the attitude matrix of the first frame rotor flow field image are spliced ​​together to obtain the splicing features;

[0089] The rotor optical flow field is determined based on the splicing characteristics.

[0090] In one embodiment of this application, the data processing module is further configured to:

[0091] The spliced ​​features are upsampled using transposed convolution to obtain the rotor optical flow field.

[0092] In one embodiment of this application, the rotor flow field determination device includes:

[0093] The model training module is used to train the optical flow field prediction model with the objective of minimizing a target loss function, thereby obtaining a pre-trained optical flow field prediction model. The target loss function includes:

[0094] ;

[0095] ;

[0096] ;

[0097] in, Let be the target loss function. As the first weight, To activate loss, As the second weight, For the reconstruction loss, i represents other object categories in the two frames of rotor flow field images, and t represents the target object category in the two frames of rotor flow field images, where the target category is floating particles. The activation probability of the target category. represents the activation probability of other categories, and m is the value that controls the difference between activation probabilities. This is the first frame of the rotor flow field image. The image is obtained by upsampling and reconstructing the structure embedding features corresponding to the first frame rotor flow field image.

[0098] It should be noted that the rotor flow field determination device and the rotor flow field determination method provided in the above embodiments belong to the same concept. The specific operation methods of each module and unit have been described in detail in the method embodiments and will not be repeated here. In practical applications, the rotor flow field determination device provided in the above embodiments can be assigned to different functional modules as needed, that is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above. This is not a limitation here.

[0099] Embodiments of this application also provide an electronic device, including: one or more processors; and a storage device for storing one or more programs, which, when executed by one or more processors, cause the electronic device to implement the rotor flow field determination method provided in the above embodiments.

[0100] Another aspect of this application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a computer's processor, causes the computer to perform the rotor flow field determination method provided in the various embodiments described above. This computer-readable storage medium may be included in the electronic device described in the above embodiments, or it may exist independently and not assembled into the electronic device.

[0101] Another aspect of this application provides a computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the rotor flow field determination method provided in the various embodiments described above.

[0102] In the embodiments of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. The terms "comprising" and "including" as used throughout the specification and claims are open-ended terms and should therefore be interpreted as "comprising but not limited to".

[0103] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A method for determining the flow field of a rotor, characterized in that, include: Acquire two adjacent frames of rotor flow field images; The two adjacent rotor flow field images are the first rotor flow field image and the second rotor flow field image, respectively. The second rotor flow field image is the image following the first rotor flow field image. Based on two adjacent frames of rotor flow field images and a pre-trained optical flow field prediction model, the rotor optical flow field is obtained. The pre-trained optical flow field prediction model is used to extract features from each frame of rotor flow field image using a capsule layer to obtain an attitude matrix. Based on the attitude matrix, structural embedding features are determined. The attitude matrix represents the geometric state of particulate matter features, and the structural embedding features represent the attitude information of particulate matter and other objects. The structural embedding features of the second frame of rotor flow field image are subtracted from the structural embedding features of the first frame of rotor flow field image to obtain motion embedding information. The rotor optical flow field is determined based on the motion embedding information; The method of extracting features from the rotor flow field image using capsule layers to obtain the attitude matrix includes: Basic features are extracted from the rotor flow field image to obtain the basic features; The first capsule layer is used to extract the first feature from the basic features to obtain the first capsule feature; The second capsule layer is used to extract the second feature from the features of the first capsule to obtain the second capsule feature. The features of the second capsule are classified to obtain the pose matrix; The determination of the rotor optical flow field based on the motion embedding information includes: The motion embedding information and the attitude matrix of the first frame rotor flow field image are spliced ​​together to obtain the splicing features; The rotor optical flow field is determined based on the splicing characteristics.

2. The rotor flow field determination method according to claim 1, characterized in that, The basic features extracted from the rotor flow field image include: The rotor flow field image is input into a convolutional layer for basic feature extraction to obtain the basic features.

3. The method for determining the rotor flow field according to claim 1, characterized in that, The determination of the rotor optical flow field based on the splicing features includes: The spliced ​​features are upsampled using transposed convolution to obtain the rotor optical flow field.

4. The method for determining the rotor flow field according to claim 1, characterized in that, Before obtaining the rotor optical flow field based on two adjacent frames of rotor flow field images and a pre-trained optical flow field prediction model, the method further includes: The optical flow field prediction model is trained with the objective of minimizing the objective loss function, resulting in a pre-trained optical flow field prediction model. The objective loss function includes: ； ； ； in, Let be the target loss function. As the first weight, To activate loss, As the second weight, For the reconstruction loss, i represents other object categories in the two frames of rotor flow field images, and t represents the target object category in the two frames of rotor flow field images, where the target category is floating particles. The activation probability of the target category. represents the activation probability of other categories, and m is the value that controls the difference between activation probabilities. This is the first frame of the rotor flow field image. The image is obtained by upsampling and reconstructing the structure embedding features corresponding to the first frame rotor flow field image.

5. A rotor flow field determination device, characterized in that, include: The data acquisition module is used to acquire two adjacent frames of rotor flow field images; The two adjacent rotor flow field images are the first rotor flow field image and the second rotor flow field image, respectively. The second rotor flow field image is the image following the first rotor flow field image. The data processing module is used to obtain the rotor optical flow field based on two adjacent frames of rotor flow field images and a pre-trained optical flow field prediction model. The pre-trained optical flow field prediction model is used to extract features from each frame of rotor flow field image using a capsule layer to obtain an attitude matrix. Based on the attitude matrix, structural embedding features are determined. The attitude matrix represents the geometric state of particulate matter features, and the structural embedding features represent the attitude information of particulate matter and other objects. The structural embedding features of the second frame of rotor flow field image are subtracted from the structural embedding features of the first frame of rotor flow field image to obtain motion embedding information. The rotor optical flow field is determined based on the motion embedding information; The data processing module also includes: The first processing unit is used to extract basic features from the rotor flow field image to obtain basic features; The second processing unit is used to perform first feature extraction on the basic features using the first capsule layer to obtain the first capsule features; The third processing unit is used to extract second features from the first capsule features using the second capsule layer to obtain the second capsule features; The fourth processing unit is used to classify the features of the second capsule to obtain the pose matrix; The data processing module is also used for: The motion embedding information and the attitude matrix of the first frame rotor flow field image are spliced ​​together to obtain the splicing features; The rotor optical flow field is determined based on the splicing characteristics.

6. An electronic device, characterized in that, The electronic device includes: One or more processors; Memory used to store the executable program code of the processor; The processor is configured to execute the program code to implement the rotor flow field determination method as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by the computer's processor, causes the computer to perform the rotor flow field determination method according to any one of claims 1 to 4.

Images