High-resolution particle image velocity measurement method based on novel cost body construction mode

Abstract

The invention relates to a high-resolution particle image velocity measurement method based on a novel cost body construction mode. The method comprises the following steps: firstly, acquiring a particle image velocity measurement simulation data set, and preprocessing the particle image velocity measurement simulation data set; then, a particle image velocity measurement model is constructed, and the model comprises a feature extraction and enhancement module, an optical flow decomposition module, a 3D cost body construction module and a velocity field prediction module; then, training the particle image velocity measurement model by adopting the preprocessed particle image velocity measurement simulation data set; and finally, inputting a to-be-predicted particle image into the trained particle image velocity measurement model to obtain a particle image prediction velocity field. That is to say, a novel feature processing module and a novel cost body construction mode are provided, so that the model can learn more complex features, the sensitivity to small-displacement small objects is remarkably improved, and accurate positioning and tracking are facilitated. The solving efficiency is improved, and a more efficient and accurate technical means is provided for fluid motion research.

Description

Technical Field

[0001] The present application relates to the technical fields of computer vision and particle image velocimetry, and in particular to a high-resolution particle image velocimetry method based on a novel cost volume construction method. Background Art

[0002] Particle image velocimetry (PIV), a key technology for non-invasively measuring velocity components in flow fields in experimental fluid dynamics, plays an important role in numerous academic research and engineering fields. PIV technology is widely used in areas such as optimizing combustion processes in automotive engines, studying the aerodynamic performance of aircraft, and assessing wind loads on buildings. To obtain a PIV particle image, buoyant tracer particles are first added to the fluid to cause them to move with the fluid. A thin, high-power laser sheet is then used to illuminate the flow area. A camera then captures two consecutive images of the particles at microsecond intervals. Analysis and processing are then performed to determine the global velocity field of the tracer particles, thereby obtaining information on the velocity distribution of the flow field.

[0003] With the rapid development of deep learning in computer vision, velocity field estimation (optical flow estimation) has achieved significant breakthroughs. Early end-to-end convolutional networks achieved fast inference speed but limited accuracy. Subsequent networks improved performance by stacking encoder-decoder networks, but these networks suffered from large model sizes and over 160 million parameters, resulting in high hardware requirements and expensive deployment. To address this issue, a series of methods developed based on the traditional coarse-to-fine pyramid strategy have reduced model size, but their ability to capture small, fast-moving objects still needs improvement. At the same time, researchers are attempting to expand optical flow estimation technology to complex scenarios such as occlusion, low light, and fog.

[0004] For PIV velocity field estimation, early artificial neural networks limited their accuracy due to their simple structures. With increasing research, various deep learning-based optical flow methods have surpassed traditional algorithms, leading to the emergence of deep neural networks specifically designed for estimating dense velocity vectors from PIV images. These models, through improved network structures and the introduction of new loss functions, have enhanced estimation accuracy. However, improved models based on specific architectures face difficulties in balancing accuracy and computational efficiency. While some improved models have improved estimation accuracy, they also suffer from significant memory consumption issues.

[0005] Research has found that as particle image resolution increases, computational costs rise significantly, and device memory usage rises exponentially, severely impacting computational efficiency. Furthermore, the accuracy of PIV particle image velocimetry still needs to be improved. Therefore, a method that can improve the accuracy of particle image velocimetry while reducing memory usage and video memory consumption is urgently needed. Summary of the Invention

[0006] Based on this, it is necessary to provide a high-resolution particle image velocimetry method based on a new cost volume construction method, which can improve the accuracy of particle image velocimetry and reduce memory occupancy and video memory consumption in order to address the above technical problems.

[0007] In a first aspect, the present application provides a high-resolution particle image velocimetry method based on a novel cost volume construction method. The method comprises: Obtain particle image velocimetry simulation data set and perform preprocessing; Constructing a particle image velocimetry model, the model includes a feature extraction and enhancement module, an optical flow decomposition module, a 3D cost volume construction module, and a velocity field prediction module; Using the preprocessed particle image velocimetry simulation data set to train the particle image velocimetry model; The particle image to be predicted is input into the trained particle image velocimetry model to obtain the particle image predicted velocity field.

[0008] Optionally, in one embodiment of the present application, the preprocessing includes scaling transformation and random horizontal flipping.

[0009] Optionally, in one embodiment of the present application, the feature extraction and enhancement module includes a feature extraction network and a feature enhancement network, the feature extraction network is constructed based on the ResNet network, and the feature enhancement network is constructed based on a multi-branch void convolution structure.

[0010] Optionally, in one embodiment of the present application, the optical flow decomposition module includes: Combine the source features and target features with the position encoding respectively and perform feature projection; Calculate the attention matrix based on the features obtained by feature projection; An optical flow decomposition result is determined based on the attention matrix and the target features.

[0011] Optionally, in one embodiment of the present application, the 3D cost volume construction module includes: Define the initial 3D cost volume; Determine a search range based on the optical flow decomposition result and the initial 3D cost volume; A new 3D cost volume is determined based on the search range.

[0012] Optionally, in one embodiment of the present application, the velocity field prediction module includes: Initialize the optical flow field, extract the optical flow features, and combine the relevant features and context features to form the input feature map; Iteratively updating the input feature map using a gated activation unit constructed based on a GRU unit to obtain a predicted optical flow image; The predicted optical flow image is upsampled and restored to the full-resolution optical flow to obtain a predicted velocity field.

[0013] Optionally, in one embodiment of the present application, the using the pre-processed particle image velocimetry simulation data set to train the particle image velocimetry model includes: The mean endpoint error metric was used to evaluate the prediction performance of the particle image velocimetry model.

[0014] In a second aspect, the present application also provides a high-resolution particle image velocimetry device based on a novel cost volume construction method. The device comprises: Data acquisition and preprocessing module, used to obtain particle image velocimetry simulation data sets and perform preprocessing; A model construction module is used to construct a particle image velocimetry model, which includes a feature extraction and enhancement module, an optical flow decomposition module, a 3D cost volume construction module, and a velocity field prediction module; A model training module is used to train the particle image velocimetry model using the preprocessed particle image velocimetry simulation data set; The particle image velocimetry module is used to input the particle image to be predicted into the trained particle image velocimetry model to obtain the particle image predicted velocity field.

[0015] In a third aspect, the present application further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program and the processor executes the steps of the method described in each of the above embodiments.

[0016] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in each of the above embodiments.

[0017] The high-resolution particle image velocimetry method based on a novel cost volume construction approach first obtains and preprocesses a particle image velocimetry simulation dataset. Next, a particle image velocimetry model is constructed, comprising a feature extraction and enhancement module, an optical flow decomposition module, a 3D cost volume construction module, and a velocity field prediction module. The model is then trained using the preprocessed particle image velocimetry simulation dataset. Finally, the particle image to be predicted is input into the trained particle image velocimetry model to obtain the predicted particle image velocity field. Specifically, a novel feature processing module is proposed, adding a feature enhancement FEM network based on multi-branch dilated convolutions to a modified 1 / 4 high-resolution feature extraction network. This allows the model to learn more complex features while supporting network depth expansion, significantly improving sensitivity to small objects with small displacements and facilitating precise localization and tracking. The multi-branch dilated convolutions learn richer local contextual features, thereby enhancing the feature representation of small objects and improving the semantic information representation of small targets. A new cost volume construction method was also proposed. By decomposing the 2D optical flow into two 1D optical flows in the horizontal and vertical directions, the original 4D pyramid is constructed into two 3D cost volumes, further reducing video memory consumption, lowering computing costs, and improving solution efficiency, providing a more efficient and accurate technical means for fluid motion research. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 FIG2 is an application environment diagram of a high-resolution particle image velocimetry method based on a novel cost volume construction method in one embodiment; Figure 2 1 is a flow chart of a high-resolution particle image velocimetry method based on a novel cost volume construction method in one embodiment; Figure 3 A schematic structural diagram of a particle image velocimetry model in one embodiment; Figure 4 Schematic diagram of the structure of a feature extraction network in one embodiment; Figure 5 Schematic diagram of the structure of a residual block in one embodiment; Figure 6 A schematic diagram of the structure of a feature enhancement network in one embodiment; Figure 7 1 is a schematic diagram of a process of optical flow decomposition in one embodiment; Figure 8 Schematic diagram of a Conv-GRU structure in one embodiment; Figure 9 A schematic diagram showing the comparison of average endpoint errors of different methods in one embodiment; Figure 101 is a structural block diagram of a high-resolution particle image velocimetry device based on a novel cost volume construction method in one embodiment; Figure 11 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0020] The high-resolution particle image velocimetry method based on a novel cost volume construction method provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown, the terminal communicates with the server through the network. The data storage system can store data that the server needs to process. The data storage system can be integrated on the server or placed on the cloud or other network servers. The terminal can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can be smart speakers, smart TVs, smart air conditioners, smart car devices, etc. Portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. The server can be implemented as a standalone server or a server cluster consisting of multiple servers.

[0021] In one embodiment, Figure 2 As shown in the figure, a high-resolution particle image velocimetry method based on a new cost volume construction method is provided. Figure 1 The following steps are used as an example to illustrate the server in the example: S201: Obtain a particle image velocimetry simulation data set and perform preprocessing.

[0022] In the examples of this application, a particle image velocimetry simulation dataset was first obtained from an internationally recognized public database for PIV algorithm benchmarks. This dataset, with its significant advantages of ideal tracer particle concentration distribution and high signal-to-noise ratio, closely approximates ideal laboratory flow conditions, providing an experimental-grade dataset for the reliability and effectiveness of network model training and evaluation results. The dataset covers the following flow types: uniform flow, a laminar benchmark; backstep flow, a classic example of flow separation and recirculation; cylinder flow, a phenomenon of vortex shedding that demonstrates complex vortex shedding dynamics; two-dimensional direct numerical simulation (DNS) turbulent motion that reveals fine turbulent coherent structures; surface quasi-geostrophic (SQG) model flows that incorporate the complexities of geophysical fluid dynamics; and public datasets such as the Johns Hopkins Turbulence Database (JHTDBs), renowned for its high-quality turbulence data and full-scale coverage. In this multivariate flow scenario, uniform flow and backstep laminar flow form the foundational validation conditions. However, the vortex shedding flow around a cylinder (Cylinder), JHTDBs turbulence, DNS turbulence, and SQG flow fields exhibit complex vortex topologies, posing significant challenges to the PIV algorithm's flow field analysis capabilities. In particular, DNS turbulence and SQG flow patterns, with their complex turbulent pulsation characteristics and unpredictable dynamic behavior, coupled with strong nonlinear coupling, provide a rigorous testing platform for evaluating the algorithm's robustness and measurement accuracy. Preprocessing operations are also performed to enhance data richness.

[0023] Specifically, in one embodiment of the present application, the preprocessing includes scaling transformation and random horizontal flipping.

[0024] In one embodiment of the present application, for image data in the training phase, a scaling transformation is introduced to enhance data diversity. The image is randomly scaled within a scaling factor range of [-0.1, 1.0], where a negative scaling factor allows the image to be slightly reduced (approximately 90% of its original size). In addition, a random horizontal flip operation is applied to further expand the training samples through random flipping with a probability of 0.5, enhancing the model's ability to recognize objects from different perspectives. These preprocessing operations effectively increase the richness of the training data, helping the model learn more generalizable feature representations, thereby improving performance in optical flow estimation tasks.

[0025] S203: Constructing a particle image velocimetry model, which includes a feature extraction and enhancement module, an optical flow decomposition module, a 3D cost volume construction module, and a velocity field prediction module.

[0026] In the embodiment of the present application, a convolutional neural network model based on a novel feature processing network and a cost body construction method is constructed, namely, a particle image velocimetry model. Figure 3As shown in the figure, the model includes a feature extraction and enhancement module, an optical flow decomposition module, a 3D cost volume construction module, and a velocity field prediction module. Specifically, the feature processing module is used to extract and enhance features, and finally the source features and target features of the image pair are extracted. A 1 / 4 high-resolution feature extraction network is used in the extraction network and a residual module is introduced to improve model performance and enhance high-density particle computing capabilities to handle complex flow field motion. In addition, due to the complexity of PIV images, the feature extraction network has limited functionality, the extracted features contain less semantic information, and the receptive field is narrow. A feature enhancement module (FEM) is added to increase feature richness through a multi-branch dilated convolutional structure, expand the network's local perception capability, and improve the semantic information expression of small targets.

[0027] Specifically, in one embodiment of the present application, the feature extraction and enhancement module includes a feature extraction network and a feature enhancement network, the feature extraction network is constructed based on the ResNet network, and the feature enhancement network is constructed based on a multi-branch void convolution structure.

[0028] In one embodiment of the present application, the feature extraction network is mainly modified based on the ResNet network, such as Figure 4 As shown in the figure, it mainly consists of five main layers (initial convolution layer, three residual block stacking layers and output convolution layer). The first main layer contains a convolution layer, a normalization layer, and an activation layer. It performs preliminary feature extraction on the input image, expands the receptive field, and normalizes the feature map to accelerate model convergence and improve the stability and generalization ability of the model. The second to fourth main layers are three layers of residual block stacking layers, each layer contains two residual blocks, as shown in the figure. Figure 5 As shown, the residual block consists of two convolutional layers, two normalization layers, and two activation layers. Residual blocks are connected via residual connections, where the output of the previous layer is added to the processed output of the current layer, and then an activation function is applied to obtain the final output of the current layer. This connection effectively solves the problems of vanishing and exploding gradients in deep neural networks, enabling the network to learn more complex features while supporting network depth expansion. The final feature extraction network outputs features at 1 / 4 resolution. This higher-resolution feature significantly improves sensitivity to small objects with small displacements, facilitating precise positioning and tracking.

[0029] After that, the extracted information enters the Feature Enhancement Module (FEM), which uses a multi-branch convolutional structure to extract multiple discriminative semantic information from the perspective of increasing feature richness. From the perspective of expanding the receptive field, dilated convolution is applied to obtain richer local context information. Figure 6As shown in Figure 1, FEM has four branches, each of which performs a 1×1 convolution operation on the input feature map to preliminarily adjust the number of channels for subsequent processing. The first branch is a residual structure that forms an equivalent map to retain key feature information of small objects. The other three branches perform cascaded standard convolution operations with kernel sizes of 1×3, 3×1, and 3×3, respectively. The two middle branches add additional dilated convolution layers so that the extracted feature maps can retain more contextual information. The formula of FEM is as follows:

[0030] in, 、 、 、 They represent standard convolution operations with kernel sizes of 1×1, 1×3, 3×1, and 3×3, respectively. represents a dilated convolution operation with a dilation rate of 5. Indicates a connection operation. Represents element-wise addition of the feature map. is the input feature map. 、 and Represents the output feature maps of the first three branches after standard convolution and dilated convolution. is the output feature map of FEM. Multi-branch dilated convolution learns richer local context features, thereby improving the feature representation ability of small objects and enhancing the semantic information expression of small targets.

[0031] In the optical flow decomposition module, a 1D self-attention mechanism is applied to the source features, and then a 1D cross-attention mechanism is applied between the source image features and the target features to factorize the 2D optical flow, allowing it to perform large displacement search on high-resolution images. Specifically, in one embodiment of the present application, the optical flow decomposition module includes: S301: Combining the source features and target features with the position codes respectively to perform feature projection.

[0032] S303: Calculate the attention matrix based on the features obtained by feature projection.

[0033] S305: Determine an optical flow decomposition result based on the attention matrix and the target features.

[0034] In one embodiment of the present application, Figure 7 As shown in , first, the feature is combined with the position encoding. The source feature and the target feature are , where H, W, and D represent height, width, and feature dimension respectively. The source feature F1 is combined with the position code P to obtain F1+P, and the target feature F2 is combined with the position code P to obtain F2+P. Feature projection is performed, and two 1×1 convolutions are used to project F1+P into the embedding space to obtain , similarly, projecting F2+P into the embedding space, we get . After that, calculate the attention matrix. and Reconstruct it into W×H×D and W×D×H, perform matrix multiplication on it to get W×H×H, and use the softmax function to normalize the last dimension to get the attention matrix Finally, at the same time Reconstruct it into W×H×D, multiply it with A to get W×H×D, and then reconstruct it to get the new feature (H×W×D), where each feature vector knows the features of the points in the same column in F2, and finally we get .

[0035] The 3D cost volume construction module significantly reduces computational costs and improves computational efficiency by constructing two 3D cost volumes with vertical attention and horizontal correlation, and horizontal attention and vertical correlation. Specifically, in one embodiment of the present application, the 3D cost volume construction module includes: S401: Define an initial 3D cost volume.

[0036] S403: Determine a search range based on the optical flow decomposition result and the initial 3D cost volume.

[0037] S405: Determine a new 3D cost volume based on the search range.

[0038] In one embodiment of the present application, first, an initial 3D cost volume is defined. , where H and W represent the height and width, R is the search radius in the horizontal direction, and the cost volume construction formula is

[0039]

[0040] in, It is a normalization factor used to prevent the dot product value from being too large.

[0041] Afterwards, the search range is determined based on the optical flow decomposition results and the initial 3D cost volume.

[0042] When i changes from 0 to H-1, Time, location The search range is the width (height) of the image in the vertical direction and extends to the maximum search radius R in the horizontal direction. Therefore, combining vertical attention and horizontal correlation, the theoretical search range is H (2R + 1).

[0043] Finally, the new 3D cost volume is determined based on the search range ,Right now

[0044] The theoretical search range here is W(2R+1). and Adding them together will give the shape The theoretical search range will also become , is the overlapping area of ​​the two cost volumes. Compared with the cost volume constructed by the traditional local window method , its theoretical search range is In comparison, the cost of constructing the cost body of the present invention is lower, and has a larger search range, further improving the computing efficiency.

[0045] Finally, in the velocity field prediction module, an iterative update is performed on the gate control unit to obtain a predicted optical flow image, and the predicted optical flow image is upsampled to restore it to full resolution and obtain a velocity field. Specifically, in one embodiment of the present application, the velocity field prediction module includes: S501: Initialize the optical flow field, extract optical flow features, and combine related features and context features to form an input feature map.

[0046] S503: Iteratively update the input feature map using a gated activation unit constructed based on the GRU unit to obtain a predicted optical flow image.

[0047] S505: Up-sampling the predicted optical flow image to restore it to full-resolution optical flow to obtain a predicted velocity field.

[0048] In one embodiment of the present application, first, the optical flow field is initialized to 0, that is, , based on the current optical flow estimation, relevant features are retrieved from the relevant pyramid and input into two convolutional layers for processing to extract effective relevant features. It applies two convolutional layers to generate optical flow features. In addition, it directly introduces the input from the context network, concatenates the above-mentioned related features, optical flow features, and context features to form an input feature map. This multi-source feature fusion method can make full use of different types of information, provide comprehensive input data for subsequent update operations, and enhance the model's ability to understand and process complex scenarios.

[0049] Afterwards, the gated activation unit constructed based on the GRU unit is used for iterative update operations, such as Figure 8 As shown, first, the hidden state of the previous moment is With the input feature map Perform splicing, calculate and update the gate through the convolution layer and reset gate , the formula is

[0050]

[0051] in, is the activation function, and is the corresponding weight matrix. Through these two gating signals, the model can adaptively control the degree of information update and retention, and decide the degree of update of the hidden state according to the current input and historical state.

[0052] After that, the candidate hidden states are calculated. Based on the reset gate and the previous hidden state , and the input feature map , calculate the candidate hidden state through the convolution layer , the formula is

[0053] in is the hyperbolic tangent activation function, The candidate hidden state comprehensively considers the current input and the historical information filtered by the reset gate, providing a new candidate value for the update of the hidden state.

[0054] After that, update the hidden state. According to the update gate and candidate hidden states , and the hidden state at the previous moment , calculate the updated hidden state , the formula is

[0055] This updating method can retain historical information while incorporating new information in a timely manner, allowing the hidden state to continuously adapt to changes in input data and gradually learn the feature representations required for more accurate optical flow estimation.

[0056] Finally, the hidden state of the GRU output Input to two convolutional layers to predict optical flow updates The output optical flow resolution is now 1 / 4 that of the input image. This multi-convolutional layer structure allows for further feature extraction and transformation of the hidden state, converting the information in the hidden state into specific optical flow updates, providing a key update direction for optimizing optical flow estimation.

[0057] After that, upsampling is performed. Since the optical flow output by the network has a resolution of 1 / 4, it needs to be restored to full resolution. Upsampling is achieved by taking the full-resolution optical flow of each pixel as a convex combination of its low-resolution 3×3 neighboring pixels. The specific operation involves using two convolutional layers to predict a The final high-resolution optical flow field is obtained by weighted combination of the neighborhood using the mask, and then dimension transformation and reshaping. This upsampling method can effectively utilize the information in the low-resolution optical flow and generate high-quality full-resolution optical flow through weighted combination, meeting the demand for high-resolution optical flow in practical applications. With the current optical flow estimate Add together to get the optical flow estimate for the next moment ,Right now Then, As the new current optical flow estimate, repeat the above steps until the optical flow estimation sequence converges to a fixed point Through continuous iterative updates, the model can gradually optimize the optical flow estimation results, making it more accurately reflect the actual movement of pixels in the image.

[0058] S205: Using the pre-processed particle image velocimetry simulation data set to train the particle image velocimetry model.

[0059] In the embodiment of the present application, in order to ensure fair and unbiased initialization, all network modules are initialized with randomly generated weights in accordance with the standard deep learning process. The AdamW optimizer is used in the training process. This optimizer effectively improves the convergence speed and generalization ability of the model by combining the adaptive learning rate mechanism and weight decay regularization strategy of the Adam algorithm. Model training is based on an open source dataset containing 13,000 pairs of images. In order to fully utilize the parallel computing capabilities of modern GPUs to accelerate training, all training tasks are completed on the NVIDIA GeForce RTX 4070 GPU. A balance between memory efficiency and training stability is achieved by selecting a batch size of 2, and the number of iterations is 120,000 rounds to ensure that the model can be exposed to a diverse set of samples during each iteration. The initial learning rate is set to λ=1×10⁻4 The weight decay coefficient of the AdamW optimizer is set to 0.00005 to regularize the model and suppress overfitting. In addition, gradient clipping is applied to constrain the gradient to the range [-1, 1], effectively solving the gradient explosion problem and enhancing the stability of the optimization process.

[0060] In one embodiment of the present application, the training of the particle image velocimetry model using the preprocessed particle image velocimetry simulation dataset includes: The mean endpoint error metric was used to evaluate the prediction performance of the particle image velocimetry model.

[0061] In one embodiment of the present application, the average endpoint error index is used to evaluate the prediction performance of the particle image velocimetry model. The smaller the average endpoint error index value, the smaller the error and the better the model prediction performance. Figure 9 The figure below is a schematic diagram of the comparison of the average endpoint errors of different methods in the experiment. The calculation formula of the indicator is as follows:

[0062] in, and They represent the predicted velocity components, and represent the true velocity components, represents the number of image pairs, Indicates the number of pixels in each image.

[0063] S207: Inputting the particle image to be predicted into the trained particle image velocimetry model to obtain a particle image predicted velocity field.

[0064] In the embodiment of the present application, finally, the particle image to be predicted is input into the trained particle image velocimetry model, and the velocity field is predicted from the particle image.

[0065] In the high-resolution particle image velocimetry method based on the novel cost volume construction approach, a particle image velocimetry simulation dataset is first acquired and preprocessed. A particle image velocimetry model is then constructed, comprising a feature extraction and enhancement module, an optical flow decomposition module, a 3D cost volume construction module, and a velocity field prediction module. The preprocessed particle image velocimetry simulation dataset is then used to train the particle image velocimetry model. Finally, the particle image to be predicted is input into the trained particle image velocimetry model to obtain the predicted particle image velocity field. Specifically, a novel feature processing module is proposed, adding a feature enhancement FEM network based on multi-branch dilated convolutions to a modified 1 / 4 high-resolution feature extraction network. This allows the model to learn more complex features while supporting network depth expansion, significantly improving sensitivity to small objects with small displacements and facilitating precise localization and tracking. The multi-branch dilated convolutions learn richer local contextual features, thereby enhancing the feature representation of small objects and improving the semantic information representation of small targets. A new cost volume construction method was also proposed. By decomposing the 2D optical flow into two 1D optical flows in the horizontal and vertical directions, the original 4D pyramid is constructed into two 3D cost volumes, further reducing video memory consumption, lowering computing costs, and improving solution efficiency, providing a more efficient and accurate technical means for fluid motion research.

[0066] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0067] Based on the same inventive concept, embodiments of the present application also provide a high-resolution particle image velocimetry device based on a novel cost volume construction method for implementing the aforementioned high-resolution particle image velocimetry method based on a novel cost volume construction method. The solution provided by this device is similar to the solution described in the aforementioned method. Therefore, the specific limitations of one or more embodiments of the high-resolution particle image velocimetry device based on a novel cost volume construction method provided below can be found in the above-mentioned limitations of the high-resolution particle image velocimetry method based on a novel cost volume construction method, and will not be repeated here.

[0068] In one embodiment, Figure 10 As shown, a high-resolution particle image velocimetry device 1000 based on a novel cost volume construction method is provided, comprising: a data acquisition and preprocessing module 1001, a model construction module 1003, a model training module 1005 and a particle image velocimetry module 1007, wherein: The data acquisition and preprocessing module 1001 is used to acquire the particle image velocimetry simulation data set and perform preprocessing.

[0069] The model construction module 1003 is used to construct a particle image velocimetry model, which includes a feature extraction and enhancement module, an optical flow decomposition module, a 3D cost volume construction module, and a velocity field prediction module.

[0070] The model training module 1005 is used to train the particle image velocimetry model using the pre-processed particle image velocimetry simulation data set.

[0071] The particle image velocimetry module 1007 is used to input the particle image to be predicted into the trained particle image velocimetry model to obtain the particle image predicted velocity field.

[0072] In one embodiment of the present application, the preprocessing includes scaling transformation and random horizontal flipping.

[0073] In one embodiment of the present application, the feature extraction and enhancement module includes a feature extraction network and a feature enhancement network, the feature extraction network is constructed based on the ResNet network, and the feature enhancement network is constructed based on a multi-branch void convolution structure.

[0074] In one embodiment of the present application, the optical flow decomposition module includes: Combine the source features and target features with the position encoding respectively and perform feature projection; Calculate the attention matrix based on the features obtained by feature projection; An optical flow decomposition result is determined based on the attention matrix and the target features.

[0075] In one embodiment of the present application, the 3D cost volume construction module includes: Define the initial 3D cost volume; Determine a search range based on the optical flow decomposition result and the initial 3D cost volume; A new 3D cost volume is determined based on the search range.

[0076] In one embodiment of the present application, the velocity field prediction module includes: Initialize the optical flow field, extract the optical flow features, and combine the relevant features and context features to form the input feature map; Iteratively updating the input feature map using a gated activation unit constructed based on a GRU unit to obtain a predicted optical flow image; The predicted optical flow image is upsampled and restored to the full-resolution optical flow to obtain a predicted velocity field.

[0077] In one embodiment of the present application, the training of the particle image velocimetry model using the preprocessed particle image velocimetry simulation dataset includes: The mean endpoint error metric was used to evaluate the prediction performance of the particle image velocimetry model.

[0078] Each module in the novel cost volume construction method for high-resolution particle image velocimetry can be implemented in whole or in part through software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device as hardware, or stored in a computer device memory as software, allowing the processor to call and execute the corresponding operations of each module.

[0079] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 11 As shown. The computer device includes a processor, memory, a communication interface, a display screen, and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal via wired or wireless communication. The wireless communication can be achieved via Wi-Fi, a mobile cellular network, NFC (near-field communication), or other technologies. When executed by the processor, the computer program implements a high-resolution particle image velocimetry method based on a novel cost volume construction method. The display screen of the computer device can be a liquid crystal display or an electronic ink display. The input device of the computer device can be a touch layer covering the display screen, keys, a trackball, or a touchpad provided on the computer device housing, or an external keyboard, touchpad, or mouse.

[0080] Those skilled in the art will understand that Figure 11 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0081] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps in the above method embodiments when executing the computer program.

[0082] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0083] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0084] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0085] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. In particular, any reference to memory, database, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processors involved in the various embodiments provided herein may be, but are not limited to, general-purpose processors, central processing units (CPUs), graphics processing units (GPUs), digital signal processors (DSPs), programmable logic devices (PLDs), data processing logic devices based on quantum computing, and the like.

[0086] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0087] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A high-resolution particle image velocimetry method based on a novel cost volume construction method, characterized in that: The method comprises: Obtain particle image velocimetry simulation data set and perform preprocessing; Constructing a particle image velocimetry model, the model includes a feature extraction and enhancement module, an optical flow decomposition module, a 3D cost volume construction module, and a velocity field prediction module; Using the preprocessed particle image velocimetry simulation data set to train the particle image velocimetry model; The particle image to be predicted is input into the trained particle image velocimetry model to obtain the particle image predicted velocity field.

2. The high-resolution particle image velocimetry method based on a novel cost volume construction method according to claim 1, characterized in that: The preprocessing includes scaling transformation and random horizontal flipping.

3. The high-resolution particle image velocimetry method based on a novel cost volume construction method according to claim 1, characterized in that: The feature extraction and enhancement module includes a feature extraction network and a feature enhancement network. The feature extraction network is constructed based on the ResNet network, and the feature enhancement network is constructed based on a multi-branch void convolution structure.

4. The high-resolution particle image velocimetry method based on a novel cost volume construction method according to claim 1, characterized in that: The optical flow decomposition module includes: Combine the source features and target features with the position encoding respectively and perform feature projection; Calculate the attention matrix based on the features obtained by feature projection; An optical flow decomposition result is determined based on the attention matrix and the target features.

5. The high-resolution particle image velocimetry method based on a novel cost volume construction method according to claim 4, characterized in that: The 3D cost volume construction module includes: Define the initial 3D cost volume; Determine a search range based on the optical flow decomposition result and the initial 3D cost volume; A new 3D cost volume is determined based on the search range.

6. The high-resolution particle image velocimetry method based on a novel cost volume construction method according to claim 1, characterized in that: The velocity field prediction module includes: Initialize the optical flow field, extract the optical flow features, and combine the relevant features and context features to form the input feature map; Iteratively updating the input feature map using a gated activation unit constructed based on a GRU unit to obtain a predicted optical flow image; The predicted optical flow image is upsampled and restored to the full-resolution optical flow to obtain a predicted velocity field.

7. The high-resolution particle image velocimetry method based on a novel cost volume construction method according to claim 1, characterized in that: The method of training the particle image velocimetry model using the pre-processed particle image velocimetry simulation data set includes: The mean endpoint error metric was used to evaluate the prediction performance of the particle image velocimetry model.

8. A high-resolution particle image velocimetry device based on a novel cost volume construction method, characterized in that: The device comprises: Data acquisition and preprocessing module, used to obtain particle image velocimetry simulation data sets and perform preprocessing; A model construction module is used to construct a particle image velocimetry model, which includes a feature extraction and enhancement module, an optical flow decomposition module, a 3D cost volume construction module, and a velocity field prediction module; A model training module is used to train the particle image velocimetry model using the preprocessed particle image velocimetry simulation data set; The particle image velocimetry module is used to input the particle image to be predicted into the trained particle image velocimetry model to obtain the particle image predicted velocity field.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images