A field camera system using a neural network field correction algorithm

Abstract

This invention provides a field-combining camera system employing a neural network field-combining correction algorithm, comprising a hardware system and a software system. The hardware system consists of a single lens, a beam-splitting prism, a dual-camera imaging chip, a micro-thread attitude adjuster, a synchronization trigger, and an image processing terminal. The beam-splitting prism enables beam splitting of the single lens's optical path, and the synchronization trigger controls the shooting sequence. The software system runs on the image processing terminal and uses a neural network field-combining correction algorithm that includes a target pre-training stage and a same-frame particle optimization stage to process images. This invention achieves nanosecond-level frame overlap time and sub-pixel-level image overlap accuracy, meeting the requirements of high sampling frequency and low frame overlap time for flow field velocimetry in high-speed wind tunnels, effectively solving the technical challenge of existing single-camera systems struggling to balance high frequency and low frame overlap time.

Description

Technical Field

[0001] This invention relates to the field of flow field measurement technology, and in particular to a field-combining camera system employing a neural network field-combining correction algorithm. Background Technology

[0002] PIV (Particle Image Velocimetry) and PTV (Particle Tracking Velocimetry) are mainstream non-contact flow field velocimetry technologies. Their core principle is to capture the displacement of particles in two frames within a short period of time using a camera, and then calculate the two-dimensional or three-dimensional velocity field. In low-speed fluid flow field measurement, the flow velocity is low, and the requirements for the camera's frame interval are not so strict (usually at the microsecond level or above), so ordinary double-exposure cameras can meet the requirements.

[0003] However, there are two key technical challenges in measuring the flow field in hypersonic wind tunnels: First, the flow field velocity is extremely high, requiring the camera to acquire two consecutive frames of particle images within a very short frame time. Otherwise, excessive particle displacement will lead to a decrease in the accuracy of cross-correlation or particle matching algorithms, thereby increasing the error in velocity field calculation. Second, existing single-camera systems cannot simultaneously meet the dual requirements of high shooting frequency and short frame time. When using dual cameras to acquire images separately, parallax will occur due to differences in optical paths and device installation deviations, resulting in inaccurate matching of the images from the two cameras, further affecting the accuracy of velocity measurement.

[0004] In the existing technology, there is a lack of an integrated system that can simultaneously meet the requirements of high-frequency shooting, low cross-frame time acquisition, and effectively eliminate the parallax error of dual cameras, which restricts the accuracy and reliability of flow field measurement in hypersonic wind tunnels. Summary of the Invention

[0005] The purpose of this invention is to provide a field-combining camera system employing a neural network field-combining correction algorithm, thereby solving the technical problems of existing technologies in high-speed wind tunnel flow field measurement, which struggle to simultaneously achieve high frequency and low frame intervals, and where dual-camera imaging parallax is difficult to eliminate. To achieve the above objective, this invention provides the following technical solution:

[0006] A field-combining camera system employing a neural network field-combining correction algorithm is characterized by comprising a hardware system and a software system. The hardware system includes a single lens, a beam splitter, a dual-camera imaging chip, a micro-threaded attitude adjuster, a synchronization trigger, and an image processing terminal. The beam splitter is positioned behind the optical path of the single lens to split the received light beam into a transmitted light path and a reflected light path, which are then transmitted to the dual-camera imaging chip. The micro-threaded attitude adjuster is connected to both the beam splitter and the dual-camera imaging chip. The synchronization trigger is electrically connected to the dual-camera imaging chip. The software system runs on the image processing terminal and processes the images acquired by the dual-camera imaging chip using a neural network field-combining correction algorithm. The neural network field-combining correction algorithm includes a target pre-training stage and a same-frame particle optimization stage. The target pre-training stage trains a neural network model based on target images to achieve pixel-level coordinate mapping between the dual-camera imaging chips. The same-frame particle optimization stage retrains the neural network model based on same-frame particle images to achieve sub-pixel-level coordinate mapping between the dual-camera imaging chips.

[0007] Therefore, through the integrated design of the system, the hardware system adopts a single-lens + beam splitter optical path structure to ensure that the imaging chips of the two cameras receive image information from the same optical path, laying the foundation for subsequent image matching; the two-stage neural network training of the software system realizes accurate coordinate mapping from pixel level to sub-pixel level, effectively eliminating parallax error in dual-camera imaging and solving the problem of imaging mismatch in existing dual-camera systems. At the same time, the single-lens design simplifies the optical path structure and reduces system complexity.

[0008] The synchronization trigger is used to control the shooting sequence of the dual-camera imaging chip, which includes a same-frame shooting sequence and a cross-frame shooting sequence. In the same-frame shooting sequence, the synchronization trigger controls the dual-camera imaging chip to acquire images simultaneously. In the cross-frame shooting sequence, the synchronization trigger controls the dual-camera imaging chip to acquire images sequentially. The synchronization trigger can also adjust the acquisition delay time of the dual-camera imaging chip, which is on the order of hundreds of nanoseconds.

[0009] Thus, the timing control function of the synchronous trigger enables flexible switching between the two shooting modes. The same-frame shooting timing provides homogeneous image data for the training of the neural network model, ensuring the effectiveness of the training samples; the nanosecond-level delay time of the cross-frame shooting timing meets the stringent requirements of high-speed wind fields for low cross-frame time, avoids velocity measurement errors caused by excessive particle displacement, and solves the problem that existing single cameras cannot simultaneously handle high frequency and low cross-frame time.

[0010] The micro-thread attitude adjuster is used to adjust the height and optical axis direction of the beam splitter prism, as well as the height and optical axis direction of the dual-camera imaging chip; through the adjustment, the field of view error of the target image acquired by the dual-camera imaging chip is kept at the pixel level.

[0011] Therefore, by using the fine adjustment function of the micro-thread attitude adjuster, the field of view error of the dual-camera imaging chip is controlled at the pixel level, which provides the basic conditions for the subsequent neural network model to achieve pixel-level coordinate mapping, reduces the training difficulty of the neural network, improves the training efficiency and mapping accuracy of the model, and avoids the problem of correction failure caused by excessive initial field of view error.

[0012] The target pre-training stage includes the following steps: controlling the dual-camera imaging chip to acquire checkerboard target images in the same frame shooting sequence; extracting corner points from the two acquired checkerboard target images to obtain the checkerboard corner point coordinates corresponding to each of the dual-camera imaging chips; constructing the neural network model based on the extracted checkerboard corner point coordinates, and training the neural network model, which is used to output the pixel-level coordinate mapping results between the dual-camera imaging chips.

[0013] Therefore, by standardizing the pre-training process of the target, the precise extraction of corner points is achieved using the regular features of the checkerboard target, providing high-quality training samples for the neural network model. The trained model can quickly achieve pixel-level coordinate mapping between the dual-camera imaging chips, laying the foundation for sub-pixel-level correction in the subsequent same-frame particle optimization stage, and improving the efficiency and stability of the entire correction process.

[0014] The step of extracting corner points from the checkerboard target image includes: defining an image intensity function f(x,y), where f(x,y) represents the correspondence between the pixel position (x,y) and the pixel gray value; performing a local Radon transform on the image and calculating the intensity integral of the lower edge ray at different angles α∈[0,π] near the pixel (x,y); and defining a response function f based on the local Radon transform result. c [x,y], the response function f c [x,y] represents the squared difference between the maximum and minimum values ​​of the local Radon transform for all angles α∈[0,π]; by finding the response function f c The local maximum value of [x,y] is used to determine the initial position of the corner point; Gaussian peak fitting is used to correct the initial position of the corner point with sub-pixel accuracy to obtain the final coordinates of the chessboard corner point.

[0015] Therefore, through a multi-step corner extraction process, from defining the image intensity function to performing local Radon transform, then calculating the response function and fitting Gaussian peak values, sub-pixel-level precise corner localization was achieved. Precise corner coordinates provide high-quality input and output data for the neural network model, ensuring the accuracy of pixel-level coordinate mapping and avoiding model training bias caused by corner extraction errors.

[0016] The Levenberg-Marquardt algorithm is used to train the neural network model. The neural network model includes an input layer, three hidden layers, and an output layer. The input layer has a dimension of R², and the dimensions of the three hidden layers are R¹, respectively. 0 、R¹ 5 、R¹ 0 The output layer dimension is R²; the activation function of the neural network model is the Tanh function; the input of the neural network model is the coordinates of the checkerboard corner point of any one of the camera imaging chips in the dual-camera imaging chip, and the output is the coordinates of the checkerboard corner point corresponding to the other camera imaging chip.

[0017] Therefore, by clearly defining the structure and training algorithm of the neural network model, the Levenberg-Marquardt algorithm combines the advantages of gradient descent and Gauss-Newton optimization, with fast convergence speed and insensitivity to initial parameters, and can efficiently find the local minimum of the loss function; the specific network layer dimensions and Tanh activation function adapt to the nonlinear requirements of coordinate mapping, ensuring that the model can accurately learn the coordinate mapping relationship between the dual-camera imaging chips and achieve pixel-level accurate mapping output.

[0018] The same-frame particle optimization stage includes the following steps: arranging a laser plane in the flow field experimental platform, such that the thickness center of the laser plane coincides with the target plane; seeding particles in the flow field experimental platform; controlling the dual-camera imaging chip to acquire multiple frames of particle images in the same-frame shooting sequence; preprocessing the acquired particle images, including median filtering, Laplacian Gaussian filtering, and threshold segmentation based on median background subtraction; and performing particle detection and extraction on the preprocessed particle images to obtain the particle center coordinates.

[0019] Therefore, through the preliminary preparation and image preprocessing process of the same-frame particle optimization stage, the design of the overlap between the laser plane and the target plane reduces the error caused by the plane difference, and the particle seeding ensures that there are enough tracking targets in the flow field; the multi-step image preprocessing effectively removes noise and background interference, provides clear image data for particle detection and extraction, ensures the extraction accuracy of particle center coordinates, and provides high-quality samples for the secondary training of the neural network model.

[0020] The same-frame particle optimization stage further includes the following steps: inputting the particle center coordinates of any one of the dual-camera imaging chips into the trained neural network model to obtain pre-biased particle coordinates; based on the pre-biased particle coordinates and the particle center coordinates of the other camera imaging chip, performing particle matching using the nearest neighbor matching method, which determines the matched particle pairs by calculating Euclidean distance to obtain the coordinates of the matched particle pairs; and using the coordinates of the matched particle pairs to perform secondary training on the neural network model, with the secondary-trained neural network model used to output the sub-pixel-level coordinate mapping results between the dual-camera imaging chips in the laser plane.

[0021] Therefore, through the process of pre-biasing, particle matching, and secondary training, the pre-biasing step reduces the initial difference in particle coordinates and improves the efficiency and accuracy of nearest neighbor matching; the nearest neighbor matching method based on Euclidean distance can accurately find corresponding particle pairs, providing accurate samples for the model's secondary training; the model after secondary training further optimizes the mapping relationship, achieves sub-pixel level coordinate mapping, eliminates the errors caused by the non-coincidence of the target plane and the laser plane and the thickness of the laser plane, and significantly improves the image correction accuracy.

[0022] The software system's image processing via a neural network field correction algorithm further includes a cross-frame velocity measurement stage, which comprises the following steps: controlling the dual-camera imaging chip to acquire particle images of the flow field in the cross-frame shooting sequence, wherein the first camera imaging chip in the dual-camera imaging chip acquires the previous frame particle image, and the second camera imaging chip acquires the next frame particle image; performing particle detection and extraction on the previous and next frame particle images respectively to obtain the particle center coordinates of the previous and next frame particle images; and inputting the particle center coordinates of the next frame particle image into the neural network model after secondary training to obtain the corrected particle coordinates.

[0023] Therefore, by using the image acquisition and correction process in the cross-frame velocity measurement stage, two consecutive frames of particle images are obtained through the cross-frame shooting sequence of the dual-camera imaging chip. Combined with the high-precision neural network model after secondary training, the particle coordinates of the next frame are corrected, ensuring the accurate correspondence of particle coordinates in the two frames. This provides a reliable data foundation for subsequent velocity field calculations and avoids velocity measurement errors caused by image mismatch.

[0024] The cross-frame velocity measurement stage further includes the following steps: calculating the particle displacement within the cross-frame time based on the particle center coordinates of the previous frame particle image and the corrected particle coordinates; obtaining the cross-frame time controlled by the synchronization trigger, and calculating the velocity field of the flow field in combination with the particle displacement, wherein the velocity field is calculated as velocity = particle displacement / cross-frame time.

[0025] Therefore, by establishing a clear velocity field calculation process, and based on the corrected particle coordinates and precise frame intervals, a simple and direct velocity calculation formula is employed to ensure the accuracy of the flow field velocity calculation. This process transforms the preliminary image acquisition and correction results into the final velocity measurement results, fully covering the entire process of hypersonic wind tunnel flow field measurement, and realizing an integrated solution from image acquisition to velocity field output. Attached Figure Description

[0026] Figure 1 A schematic diagram of a field-combining camera system employing a neural network field-combining correction algorithm, according to an embodiment of the present invention, is shown.

[0027] Figure 2 The network structure of a neural network model of a field-combining camera system employing a neural network field-combining correction algorithm, according to an embodiment of the present invention, is shown.

[0028] Figure 3 The diagram shows a velocity field distribution obtained by a field-combining camera system employing a neural network field-combining correction algorithm, according to an embodiment of the present invention. Detailed Implementation

[0029] Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the following description, the same reference numerals are used for the same parts, and repeated descriptions are omitted. Furthermore, the drawings are merely schematic diagrams, and the proportions of the parts or the shapes of the parts may differ from the actual figures.

[0030] This invention relates to a field-combining camera system employing a neural network field-combining correction algorithm. Through optical path optimization design in the hardware system and neural network correction algorithm in the software system, it achieves high-frequency, low-frame-time measurement of hypersonic wind tunnel flow fields. The hardware system is responsible for accurate image acquisition, while the software system is responsible for image correction and velocity field calculation. The two work together to ensure measurement accuracy.

[0031] In this embodiment, the hardware system includes a single lens and a beam splitter ( Figure 1 The system comprises a dual-camera imaging chip, a micro-threaded attitude adjuster, a synchronization trigger, and an image processing terminal. The functions of each component are as follows:

[0032] Single lens: Used to receive light signals in the flow field, providing a basic optical path for subsequent imaging and ensuring that light can accurately enter the beam splitter;

[0033] Beam splitter: Located behind the optical path of a single lens, its core function is to split the optical path received by the single lens into a transmitted optical path and a reflected optical path, so that the image information of the same optical path can be transmitted to the imaging chips of two cameras respectively, thus realizing the acquisition of images from the same source.

[0034] Dual-camera imaging chip: used to receive the optical path signal after beam splitting by the beam splitter and to form an image. The two chips correspond to the transmitted optical path and the reflected optical path respectively, ensuring that images can be acquired synchronously or asynchronously.

[0035] Micro-thread attitude adjuster: It is connected to the beam splitter and the dual-camera imaging chip respectively, and can finely adjust the height and optical axis direction of the beam splitter and the height and optical axis direction of the dual-camera imaging chip to achieve preliminary correction of field error.

[0036] Synchronization trigger: Electrically connected to the dual-camera imaging chip, used to precisely control the shooting sequence of the two chips, including two modes: same-frame shooting and cross-frame shooting;

[0037] Image processing terminal: Used to run software systems, store acquired image data, and execute neural network field correction algorithms and velocity field calculations.

[0038] In this embodiment, the hardware system setup steps are as follows:

[0039] Step 1: Position the calibration target at the laser alignment point, ensuring that the target can be clearly captured by the lens to provide a standard image for subsequent calibration;

[0040] Step 2: Set up a single lens directly facing the target, adjust the lens angle to ensure that the light path enters the lens perpendicularly to the target, and ensure the clarity and accuracy of the image;

[0041] Step 3: Place a beam splitter directly behind the single lens, and install a micro-thread attitude adjuster below the beam splitter to adjust the height and attitude of the beam splitter to ensure that the light path enters perpendicularly to the center of the beam splitter surface, and that the transmitted light path and reflected light path are complete, so that both camera imaging chips can receive the complete image.

[0042] Step 4: Place camera imaging chips at the exit ends of the transmission and reflection light paths of the beam splitter, respectively, ensuring that the two chips are at the same height and that the light enters the chips perpendicularly to avoid image distortion caused by installation angle deviation.

[0043] Step 5: Electrically connect the synchronization trigger to the dual-camera imaging chip, and establish a data connection between the image processing terminal and the dual-camera imaging chip and the synchronization trigger to ensure that image data can be transmitted to the terminal in real time, and that the trigger can effectively control the chip's shooting sequence.

[0044] In this embodiment, the synchronization trigger supports two shooting timing settings, as follows:

[0045] Simultaneous frame capture timing: A synchronization signal is sent through a synchronization trigger to control the imaging chips of the two cameras to start acquiring images at the same time, ensuring that the flow field images are acquired at the same moment for training the neural network model.

[0046] Cross-frame shooting timing: The acquisition delay time of the two camera imaging chips is set by a synchronous trigger, and the chips are controlled to acquire images sequentially. The delay time can be adjusted to the level of hundreds of nanoseconds to meet the requirements of high-speed wind fields for low cross-frame time, and is used for velocity measurement of actual flow fields.

[0047] In this embodiment, the software system runs on an image processing terminal. Its core is a neural network field correction algorithm, which includes a target pre-training stage, a same-frame particle optimization stage, and a cross-frame velocity measurement stage. The process for each stage is as follows:

[0048] The core objective of the target pre-training phase is to train a neural network model based on checkerboard target images to achieve pixel-level coordinate mapping between the dual-camera imaging chips. The specific steps are as follows:

[0049] Step 1: Control the dual-camera imaging chip to acquire chessboard target images in the same frame sequence. During the acquisition process, ensure that the target is stationary and the imaging brightness is uniform to avoid image quality degradation caused by ambient light interference.

[0050] Step 2: Extract corner points from the two acquired chessboard target images. The specific process is as follows:

[0051] Define an image intensity function f(x,y), which represents the correspondence between the position (x,y) of a pixel in the image and the gray value of that pixel, providing basic data for subsequent image processing.

[0052] Perform local Radon transform R on the image flocal [x,y,α], calculate the intensity integral of the lower edge ray at different angles α∈[0,π] near the pixel (x,y). The formula for the local Radon transform is:

[0053]

[0054] Where m is the expected size of the checkerboard corner point, f[x,y] is the image intensity function at point (x,y), and k represents the position along the ray.

[0055] The response function f is defined based on the local Radon transform result. c [x,y], this function is the squared difference between the maximum and minimum values ​​of the local Radon transform for all angles α∈[0,π], used to evaluate the probability of each pixel being a corner point of the checkerboard grid. The response function formula is:

[0056]

[0057] By finding the response function f c The local maxima of [x,y] are used to determine the initial positions of the corner points. These local maxima points exhibit significant intensity differences in different directions around them, which is consistent with the characteristics of checkerboard corner points.

[0058] Gaussian peak fitting is used to correct the initially determined corner positions with sub-pixel accuracy, further improving the accuracy of the corner coordinates and ensuring the accuracy of subsequent model training.

[0059] Step 3: Construct a neural network model based on the extracted chessboard corner coordinates (e.g., Figure 2 The model takes the coordinates of the checkerboard corner points of one of the camera imaging chips as input and outputs the coordinates of the checkerboard corner points of the other camera imaging chip.

[0060] Step 4: Train the neural network model using the Levenberg-Marquardt (LM) algorithm. This algorithm combines the features of gradient descent and Gauss-Newton optimization. By introducing adjustable parameters to balance the trade-off between the two, it achieves fast convergence and is insensitive to initial parameters.

[0061] The structure of the neural network model is as follows: the input layer has a dimension of R² (two-dimensional coordinates of the corresponding corner points), and the dimensions of the three hidden layers are R¹ respectively. 0 、R¹ 5 、R¹ 0 The output layer has an R² dimension and uses the Tanh function as the activation function.

[0062] Define the error function e, and calculate it using the following formula:

[0063]

[0064] Where g( )= The parameters of the neural network are: The Euclidean distance between the predicted coordinates and the true coordinates is N, where N is the number of training samples.

[0065] Initialize the network and training parameters, and calculate the initial error E0.

[0066] Calculate the Jacobian matrix J:

[0067]

[0068] The matrix elements are the partial derivatives of the error function with respect to each network parameter.

[0069] Calculate the iterative value of the parameters θ, the formula is:

[0070]

[0071] Where μ is the damping factor.

[0072] Update the network parameters, calculate the new error E1, and compare E0 with E1: if E0 > E1, accept this iteration, update the parameters and decrease μ; if E0 ≤ E1, reject this iteration, increase μ and recalculate. θ.

[0073] The iterative process is repeated until the error meets the preset threshold. After training, a pre-biased neural network model is obtained, realizing pixel-level coordinate mapping between the dual-camera imaging chips.

[0074] The core objective of the same-frame particle optimization stage is to train a neural network model a second time based on the same-frame particle image to achieve sub-pixel level coordinate mapping. The specific steps are as follows:

[0075] Step 1: Arrange the laser plane in the flow field experimental platform, adjust the angle and position of the laser generator so that the thickness center of the laser plane coincides with the target plane, reduce the error caused by the plane difference, and optimize the particle matching effect;

[0076] Step 2: Remove the target, turn on the particle generator, and seed particles into the flow field to ensure that the particles uniformly fill the camera's field of view, providing sufficient targets for particle detection and matching;

[0077] Step 3: Control the dual-camera imaging chip to acquire multiple frames of particle images in the same frame sequence. The number of frames acquired is adjusted according to the flow field stability to ensure sufficient training samples.

[0078] Step 4: Preprocess the acquired particle images, specifically including:

[0079] Median filtering: Removes salt-and-pepper noise from an image while preserving particle edge information;

[0080] Laplacian Gaussian filter: Further removes noise and enhances the contrast between particles and the background. The default filter size is 3x3 pixels, which can be adjusted according to the noise level of the image.

[0081] Thresholding segmentation: The image is divided into background image B based on median background subtraction. t and foreground image F t The calculation formula is:

[0082]

[0083]

[0084]

[0085] Where B0 is the initial background image, obtained based on the initial few frames, β is the background adaptation speed function, which controls the background image update speed, and sign is the sign function, indicating the direction of the difference between the current frame and the background;

[0086] Step 5: Perform particle detection and extraction on the preprocessed particle image to obtain the particle center coordinates:

[0087] By combining the Differential Gaussian (DoG) and Laplacian Gaussian (LoG) methods to filter the image, an image F with a strong response to particle position is obtained. t ', The high numerical value domain in the image is the particle pixel region;

[0088] Through F t The local maximum value is found to determine the center position of the particle. An intensity threshold is set, and the maximum value below the threshold is ignored to avoid false detection.

[0089] By fitting a quadratic or Gaussian function to the center position, the sub-pixel level particle center coordinates are calculated. Assuming a local maximum is detected at pixel (i, j), the sub-pixel level coordinates are:

[0090]

[0091]

[0092] Step 6: Input the particle center coordinates of one of the camera imaging chips into the neural network model obtained in the target pre-training stage to obtain the pre-biased particle coordinates and reduce the initial difference in particle coordinates.

[0093] Step 7: Perform particle matching using the nearest neighbor matching method and calculate the pre-biased particle coordinates. x i2,t , y i2,t), arbitrarily select a particle's coordinates and calculate the coordinates of all particles in the same frame as the camera ( x j1,t , y j1,t The Euclidean distance between and is given by the formula:

[0094]

[0095] in These are the pre-biased particle coordinates. Given the particle center coordinates of another camera, identify the nearest particle pair as the matching particle pair and obtain the coordinates of the matching particle pair;

[0096] Step 8: Use the matched particle pair coordinates to perform secondary training on the neural network model. The training algorithm and network structure are consistent with the target pre-training stage. After secondary training, the model can output sub-pixel level coordinate mapping results between the dual-camera imaging chips in the laser plane.

[0097] In the cross-frame velocity measurement phase, the core objective is to acquire cross-frame particle images, correct the images based on a trained neural network model, and calculate the flow velocity field. The specific steps are as follows:

[0098] Step 1: Control the dual-camera imaging chip to capture particle images of the flow field in a time-series manner across frames. The first camera imaging chip captures the previous frame particle image A, and the second camera imaging chip captures the next frame particle image B. The cross-frame time controlled by the synchronous trigger is on the order of hundreds of nanoseconds.

[0099] Step 2: Perform particle detection and extraction on the previous frame image A and the next frame image B respectively. The extraction process is the same as steps 4-5 in the same frame particle optimization stage, and obtain the particle center coordinates of each of the two frames.

[0100] Step 3: Input the particle center coordinates of the next frame image B into the neural network model after secondary training to obtain the corrected particle coordinates, ensuring the accurate correspondence of particle coordinates in the two frames.

[0101] Step 4: Based on the particle center coordinates of the previous frame image A and the corrected particle coordinates, calculate the particle displacement over the time span.

[0102] Step 5: Obtain the frame transition time controlled by the synchronization trigger. Calculate the velocity field of the flow field using the formula: Velocity = Particle Displacement / Frame Transition Time (e.g., ...). Figure 3 (As shown).

[0103] The embodiments described above do not constitute a limitation on the scope of protection of this technical solution. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the above embodiments should be included within the scope of protection of this technical solution.

Claims

1. A field-combining camera system employing a neural network field-combining correction algorithm, characterized in that, Including hardware systems and software systems; The hardware system includes a single lens, a beam splitter, a dual-camera imaging chip, a micro-threaded attitude adjuster, a synchronization trigger, and an image processing terminal. The beam splitter is positioned behind the optical path of the single lens and is used to split the light received by the single lens into a transmitted light path and a reflected light path, which are then transmitted to the dual-camera imaging chip, respectively. The micro-threaded attitude adjuster is connected to both the beam splitter and the dual-camera imaging chip. The synchronization trigger is electrically connected to the dual-camera imaging chip. The software system runs on the image processing terminal and is used to process the images acquired by the dual-camera imaging chips through a neural network field correction algorithm. The neural network field correction algorithm includes a target pre-training stage and a same-frame particle optimization stage. The target pre-training stage is used to train a neural network model based on the target image to achieve pixel-level coordinate mapping between the dual-camera imaging chips. The same-frame particle optimization stage is used to retrain the neural network model based on the same-frame particle image to achieve sub-pixel-level coordinate mapping between the dual-camera imaging chips.

2. The system according to claim 1, characterized in that, The synchronization trigger is used to control the shooting sequence of the dual-camera imaging chip, which includes intra-frame shooting sequence and cross-frame shooting sequence; under the intra-frame shooting sequence, the synchronization trigger controls the dual-camera imaging chip to acquire images simultaneously. In the cross-frame shooting sequence, the synchronization trigger controls the dual-camera imaging chip to acquire images sequentially, and the synchronization trigger can adjust the acquisition delay time of the dual-camera imaging chip, which is on the order of hundreds of nanoseconds.

3. The system according to claim 1, characterized in that, The micro-thread attitude adjuster is used to adjust the height and optical axis direction of the beam splitter prism, as well as the height and optical axis direction of the dual-camera imaging chip; through the adjustment, the field of view error of the target image acquired by the dual-camera imaging chip is kept at the pixel level.

4. The system according to claim 2, characterized in that, The target pre-training phase includes the following steps: The dual-camera imaging chip is controlled to acquire chessboard target images in the same frame shooting sequence; corner points are extracted from the two acquired chessboard target images to obtain the chessboard corner point coordinates corresponding to each of the dual-camera imaging chips; the neural network model is constructed based on the extracted chessboard corner point coordinates and trained; the neural network model is used to output the pixel-level coordinate mapping results between the dual-camera imaging chips.

5. The system according to claim 4, characterized in that, The step of extracting corner points from the chessboard target image includes: Define an image intensity function f(x,y), which represents the correspondence between the pixel position (x,y) and the pixel gray value in the image; Perform a local Radon transform on the image and calculate the intensity integral of the lower ray at different angles α∈[0,π] near the pixel (x,y); The response function f is defined based on the local Radon transform result. c [x,y], the response function f c [x,y] is the squared difference between the maximum and minimum values ​​of the local Radon transform for all angles α∈[0,π]. By finding the response function f c Find the local maximum value of [x,y] to determine the initial position of the corner point; Gaussian peak fitting is used to correct the initial position of the corner point with sub-pixel accuracy to obtain the final coordinates of the chessboard corner point.

6. The system according to claim 4, characterized in that, The neural network model is trained using the Levenberg-Marquardt algorithm; the neural network model includes an input layer, three hidden layers, and an output layer, wherein the input layer has a dimension of [missing value]. The dimensions of the three hidden layers are as follows: , , The output layer dimension is The activation function of the neural network model is the Tanh function; the input of the neural network model is the coordinates of the checkerboard corner point of any one of the camera imaging chips in the dual-camera imaging chip, and the output is the coordinates of the checkerboard corner point corresponding to the other camera imaging chip.

7. The system according to claim 2, characterized in that, The same-frame particle optimization stage includes the following steps: A laser plane is arranged in the flow field experimental platform, such that the center of the thickness of the laser plane coincides with the target plane; Particles were seeded in the flow field experimental platform; The dual-camera imaging chip is controlled to acquire multiple frames of particle images in the same frame shooting sequence; The acquired particle images are preprocessed, including median filtering, Laplacian Gaussian filtering, and threshold segmentation based on median background subtraction. The preprocessed particle image is subjected to particle detection and extraction to obtain the particle center coordinates.

8. The system according to claim 7, characterized in that, The same-frame particle optimization stage also includes the following steps: The particle center coordinates of any one of the dual-camera imaging chips are input into the trained neural network model to obtain the pre-biased particle coordinates. Based on the pre-biased particle coordinates and the particle center coordinates of another camera imaging chip, the nearest neighbor matching method is used for particle matching. The nearest neighbor matching method determines the matched particle pairs by calculating the Euclidean distance and obtains the coordinates of the matched particle pairs. The neural network model is trained a second time using the coordinates of the matched particle pairs. The neural network model after the second training is used to output the sub-pixel level coordinate mapping result between the dual-camera imaging chips in the laser plane.

9. The system according to claim 8, characterized in that, The process of image processing by the software system using the neural network field correction algorithm also includes a cross-frame speed measurement stage, which includes the following steps: The dual-camera imaging chip is controlled to acquire particle images of the flow field in the cross-frame shooting sequence, wherein the first camera imaging chip in the dual-camera imaging chip acquires the previous frame particle image and the second camera imaging chip acquires the next frame particle image. Particle detection and extraction are performed on the previous frame particle image and the next frame particle image respectively to obtain the particle center coordinates of the previous frame particle image and the particle center coordinates of the next frame particle image. The particle center coordinates of the next frame of particle image are input into the neural network model after secondary training to obtain the corrected particle coordinates.

10. The system according to claim 9, characterized in that, The cross-frame speed measurement phase also includes the following steps: Based on the particle center coordinates of the previous frame particle image and the corrected particle coordinates, the displacement of the particle during the cross-frame time is calculated. The cross-frame time controlled by the synchronous trigger is obtained, and the velocity field of the flow field is calculated by combining the displacement of the particle. The velocity field is calculated as velocity = particle displacement / cross-frame time.

Images