CN-121499852-B - Field-in-place camera system adopting neural network complex field correction algorithm

CN121499852BCN 121499852 BCN121499852 BCN 121499852BCN-121499852-B

Abstract

The invention provides a field-applied camera system adopting a neural network complex field correction algorithm, which comprises a hardware system and a software system. The hardware system consists of a single lens, a beam splitting prism, a double-camera imaging chip, a micro-thread attitude regulator, a synchronous trigger and an image processing terminal, wherein the beam splitting of a single lens light path is realized through the beam splitting prism, the synchronous trigger controls shooting time sequence, and the software system runs on the image processing terminal and processes images by adopting a neural network complexing field correction algorithm comprising a target pre-training stage and a same-frame particle optimization stage. The invention realizes the overlapping precision of hundred nanosecond-level frame-crossing time and sub-pixel-level images, meets the requirement of flow field speed measurement in a high overspeed wind tunnel on high sampling frequency and low frame-crossing time, and effectively solves the technical problem that the existing single-camera system is difficult to consider both high frequency and low frame-crossing time.

Inventors

PAN LI
ZHONG RUI
WANG SHAOFEI
WANG JINJUN
LIU YUNPENG

Assignees

北京航空航天大学宁波创新研究院
北京航空航天大学

Dates

Publication Date: 20260505
Application Date: 20260113

Claims (10)

1. A field-in-place camera system adopting a neural network complex field correction algorithm is characterized in that, Including hardware systems and software systems; The hardware system comprises a single lens, a beam splitting prism, a double-camera imaging chip, a micro-thread posture regulator, a synchronous trigger and an image processing terminal, wherein the beam splitting prism is arranged behind a light path of the single lens and is used for splitting a light path received by the single lens into a transmission light path and a reflection light path, and the transmission light path and the reflection light path are respectively transmitted to the double-camera imaging chip; The software system is operated at the image processing terminal and is used for processing images acquired by the two-camera imaging chips through a neural network complexing field correction algorithm, the neural network complexing field correction algorithm comprises a target pre-training stage and a co-frame particle optimization stage, the target pre-training stage is used for training a neural network model based on target images to achieve pixel-level coordinate mapping between the two-camera imaging chips, and the co-frame particle optimization stage is used for training the neural network model based on co-frame particle images for the second time to achieve sub-pixel-level coordinate mapping between the two-camera imaging chips.
2. The system of claim 1, wherein the synchronization trigger is configured to control a capturing timing of the dual-camera imaging chip, the capturing timing includes a co-frame capturing timing and a cross-frame capturing timing, the synchronization trigger is configured to control the dual-camera imaging chip to simultaneously capture images at the co-frame capturing timing, the synchronization trigger is configured to control the dual-camera imaging chip to sequentially capture images at the cross-frame capturing timing, and the synchronization trigger is configured to adjust a capture delay time of the dual-camera imaging chip, the capture delay time being in a hundred nanoseconds.
3. The system of claim 1, wherein the micro-threaded pose adjuster is configured to adjust a height and an optical axis direction of the beam splitter prism and a height and an optical axis direction of the dual-camera imaging chip, and wherein the adjustment is configured to cause a field of view error of a target image acquired by the dual-camera imaging chip to be at a pixel level.
4. The system of claim 2, wherein the target pre-training phase comprises the steps of: the method comprises the steps of controlling a dual-camera imaging chip to acquire checkerboard target images at the same frame shooting time sequence, respectively extracting angular points of the acquired two checkerboard target images to obtain corresponding checkerboard angular point coordinates of the dual-camera imaging chip, constructing a neural network model based on the extracted checkerboard angular point coordinates, training the neural network model, and outputting pixel-level coordinate mapping results among the dual-camera imaging chips by the neural network model.
5. The system of claim 4, wherein the step of corner extraction of the checkerboard target image comprises: Defining an image intensity function f (x, y), wherein the image intensity function f (x, y) represents the corresponding relation between the pixel point position (x, y) on the image and the pixel gray value; Carrying out local Radon transformation on the image, and calculating the intensity integral of the lower edge rays of different angles alpha E [0, pi ] near the pixel point (x, y); Defining a response function f c [ x, y ] based on the local Radon transformation result, wherein the response function f c [ x, y ] is the square difference between the maximum value and the minimum value of the local Radon transformation under all angles alpha E [0, pi ]; Determining the preliminary position of the corner point by searching the local maximum value of the response function f c [ x, y ]; And carrying out sub-pixel level precision correction on the initial positions of the angular points by adopting Gaussian peak fitting to obtain final angular point coordinates of the checkerboard.
6. The system of claim 4, wherein training the neural network model uses a Levenberg-Marquardt algorithm, the neural network model comprising an input layer, a triple hidden layer, and an output layer, the input layer being of a dimension of The dimensions of the triple hidden layers are in turn 、、 The dimension of the output layer is The activation function of the neural network model is a Tanh function, the input of the neural network model is the checkerboard angular point coordinate of any one camera imaging chip in the two-camera imaging chip, and the output is the checkerboard angular point coordinate corresponding to the other camera imaging chip.
7. The system according to claim 2, wherein the co-framed particle optimization phase comprises the steps of: Arranging a laser plane in a flow field experiment platform, so that the thickness center of the laser plane coincides with a target plane; sowing particles in the flow field experiment platform; Controlling the dual-camera imaging chip to acquire multi-frame particle images at the same frame shooting time sequence; Preprocessing the acquired particle image, wherein the preprocessing comprises median filtering, laplacian Gaussian filtering and threshold segmentation based on median background subtraction; And carrying out particle detection and extraction on the pretreated particle image to obtain a particle center coordinate.
8. The system of claim 7, wherein the co-frame particle optimization stage further comprises the steps of: inputting the particle center coordinates of any one of the two 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 by adopting a nearest neighbor matching method, and determining matched particle pairs by calculating Euclidean distances by the nearest neighbor matching method to obtain matched particle pair coordinates; And carrying out secondary training on the neural network model by using the matched particles, wherein the neural network model after the secondary training is used for outputting a sub-pixel level coordinate mapping result between the two camera imaging chips in a laser plane.
9. The system of claim 8, wherein the software system processes the image via a neural network complex field correction algorithm further comprises a cross-frame velocimetry phase comprising the steps of: Controlling the two-camera imaging chip to acquire particle images of a flow field at the cross-frame shooting time sequence, wherein a first camera imaging chip in the two-camera imaging chip acquires a previous frame of particle images and a second camera imaging chip acquires a subsequent frame of particle images; Respectively detecting and extracting particles of the previous frame of particle image and the subsequent frame of particle image to obtain the particle center coordinates of the previous frame of particle image and the particle center coordinates of the subsequent frame of particle image; and inputting the particle center coordinates of the particle image of the next frame into the neural network model after the secondary training to obtain corrected particle coordinates.
10. The system of claim 9, wherein the cross-frame velocimetry phase further comprises the steps of: calculating the displacement of particles in the cross-frame time based on the particle center coordinates of the previous frame particle image and the corrected particle coordinates; And acquiring the cross-frame time controlled by the synchronous trigger, and calculating the speed field of the flow field by combining the displacement of the particles, wherein the speed field is calculated in a mode of speed=particle displacement/cross-frame time.

Description

Field-in-place camera system adopting neural network complex field correction algorithm Technical Field The invention relates to the technical field of flow field measurement, in particular to a field-closing camera system adopting a neural network complexing field correction algorithm. Background PIV (particle image velocimetry) and PTV (particle tracking velocimetry) technologies are mainstream non-contact flow field velocimetry technologies, and the core principle of the technology is that displacement of two frames of particle images is captured in a short time through a camera, so that a two-dimensional or three-dimensional velocity field is calculated. In low-speed fluid flow field measurement, the flow speed is low, the requirements on the cross-frame time of a camera are relaxed (generally more than microsecond), and the requirements can be met by a common double-exposure camera. In a high overspeed wind tunnel flow field measurement scene, however, two key technical pain points exist, namely, the flow field flow velocity is extremely high, a camera is required to acquire two continuous frames of particle images in extremely short frame crossing time, otherwise, the particle displacement is excessively large, so that the accuracy of a cross-correlation or particle matching algorithm is reduced, further, the speed field calculation error is increased, the existing single-camera system is difficult to simultaneously meet the dual requirements of high shooting frequency and low frame crossing time, and parallax is generated due to light path difference, device installation deviation and the like when the two cameras are adopted to respectively acquire images, so that the two-camera imaging cannot be accurately matched, and further the speed measurement accuracy is influenced. In the prior art, an integrated system capable of simultaneously meeting high-frequency shooting and low-frame-crossing time acquisition and effectively eliminating parallax errors of two cameras is lacking, and accuracy and reliability of flow field measurement of a high overspeed wind tunnel are restricted. Disclosure of Invention The invention aims to provide a field-closing camera system adopting a neural network complex field correction algorithm, which solves the technical problems that the prior art is difficult to consider high frequency and low frame crossing time in the measurement of a high overspeed wind tunnel flow field, and the parallax of two camera imaging is difficult to eliminate. In order to achieve the above purpose, the present invention provides the following technical solutions: A field-closing camera system adopting a neural network complexing field correction algorithm is characterized by comprising a hardware system and a software system, wherein the hardware system comprises a single lens, a beam splitting prism, a two-phase camera imaging chip, a micro-thread posture regulator, a synchronous trigger and an image processing terminal, the beam splitting prism is arranged behind an optical path of the single lens and used for splitting an optical path received by the single lens into a transmission optical path and a reflection optical path, the transmission optical path and the reflection optical path are respectively transmitted to the two-phase camera imaging chip, the micro-thread posture regulator is respectively connected with the beam splitting prism and the two-phase camera imaging chip, the synchronous trigger is electrically connected with the two-phase camera imaging chip, the software system is operated at the image processing terminal and used for processing images acquired by the two-phase camera imaging chip through the neural network complexing field correction algorithm, the neural network complexing field correction algorithm comprises a target pre-training stage and a co-frame particle optimization stage, the target pre-training stage is used for training a neural network model based on a target image, and used for realizing pixel-level coordinate mapping among the two-phase image imaging chips and co-frame sub-training particle co-pixel mapping sub-level image sub-training stage. The two-stage neural network training of the software system realizes the accurate coordinate mapping from pixel level to sub-pixel level, effectively eliminates parallax error of double-camera imaging, solves the problem of mismatching of the imaging of the existing double-camera system, simplifies the optical path structure and reduces the complexity of the system. The synchronous trigger is used for controlling the shooting time sequence of the two-camera imaging chip, the shooting time sequence comprises a same-frame shooting time sequence and a cross-frame shooting time sequence, the synchronous trigger controls the two-camera imaging chip to simultaneously acquire images under the same-frame shooting time sequence, the synchronous trigger controls the two-camera imaging chip to sequentially acquire image