A method and system for detecting microdroplet jet parameters using multi-view single-frame stroboscopic imaging

Abstract

This invention discloses a method and system for detecting droplet parameters in microdroplet jetting using multi-view single-frame stroboscopic imaging, belonging to the field of microdroplet jetting additive manufacturing technology. It includes: organically combining single-frame multi-pulse temporal imaging, multi-view 3D reconstruction, multi-view consistency constraints, and deep learning image processing technology to simultaneously acquire the 3D spatiotemporal information (3D morphology, volume, and velocity) of the main droplet during microdroplet jetting additive manufacturing within a single exposure. Based on this, reliable synchronous detection and identification of satellite droplets is verified by achieving 3D ellipsoidal morphology constraints and independent trajectory continuity. It features simple structure, low cost, good compatibility, high measurement accuracy, strong robustness, and strong real-time performance.

Description

Technical Field

[0001] This invention belongs to the field of microdroplet jet additive manufacturing technology, specifically relating to a method and system for detecting microdroplet jet parameters using multi-view single-frame stroboscopic technology. Background Technology

[0002] Microdroplet jetting additive manufacturing technology, with its advantages of high precision, non-contact operation, and high material utilization, has been widely used in fields such as conformal electronics and flexible electronics. During microdroplet jetting, the volume, velocity, and flight attitude of the jetting droplets directly affect the geometric accuracy and functional characteristics of the final formed part. Simultaneously, the tiny satellite droplets generated alongside the main droplet can cause defects such as scattering and trailing in the printed pattern. Therefore, real-time detection of droplet morphology, especially the identification of satellite droplets, is a crucial aspect of quality control.

[0003] Currently, visual inspection is the most widely used method for detecting flying droplets. It involves capturing images of the flying droplets with a camera and then processing the images to extract parameters such as the droplet's contour, velocity, and volume. Depending on the imaging method, visual inspection is further divided into single-view and multi-view inspection. Single-view inspection is the mainstream approach, using a single camera to capture a projected image of the droplet from a single direction. This method uses stroboscopic illumination to freeze the high-speed flying droplet, obtaining a clear image before calculating the droplet's equivalent diameter and velocity on the projection plane. Single-view detection has the advantages of simple system structure and low cost, but its limitations are also quite prominent: First, a single view can only acquire two-dimensional information of the droplet on a certain projection plane, and cannot reflect the true three-dimensional shape of the droplet. When the droplet undergoes asymmetric deformation or there are multiple satellite droplets distributed in different spatial positions, single-view measurement will introduce significant errors. Second, velocity measurement under single view usually relies on multiple frame image sequences, which requires a high camera frame rate, usually requiring a high-speed camera with thousands of frames, resulting in a significant increase in hardware investment. Third, single view has difficulty in accurately distinguishing between spatially overlapping main droplets and satellite droplets. When satellite droplets are on the projection path of the main droplet, the two completely overlap in the image and cannot be effectively identified. Finally, existing single-view stroboscopic imaging technology usually adopts the method of synchronous stroboscopic acquisition of multiple frames of images, which requires a high camera frame rate, and the volume and velocity parameters need to be calculated separately through different image sequences, making it difficult to achieve synchronous high-precision measurement in a single acquisition process.

[0004] To address the limitations of single-view detection, a few studies have proposed dual-view detection schemes, which involve deploying cameras in two directions to simultaneously acquire droplet images and then fusing the dual views to obtain more comprehensive droplet information. However, existing schemes generally employ multi-frame acquisition strategies, requiring each camera to continuously acquire multiple frames to complete velocity measurements. This still places high demands on the camera frame rate (still requiring thousands of frames per second), and the droplet morphology may change between different frames in multi-frame acquisition mode, affecting measurement consistency.

[0005] Some studies have proposed a scheme that combines a single camera with an optical path deflection element and a stroboscopic light source, which can record multiple images of the same droplet at different times in a single frame. However, this scheme only has one perspective and cannot obtain the three-dimensional morphological information of the droplet, nor can it accurately locate satellite droplets distributed in space.

[0006] Furthermore, existing methods for identifying satellite droplets largely rely on morphological threshold judgments from a single viewpoint. When satellite droplets overlap with the main droplet in the projection direction, they cannot be effectively distinguished, and there is a lack of multi-view geometric consistency constraints to verify the authenticity of the satellite droplets. Simultaneously, existing droplet image processing methods primarily depend on traditional image segmentation and edge detection algorithms. When dealing with complex scenarios such as droplet overlap, occlusion, blurred boundaries, and significant scale differences between satellite and main droplets, their segmentation accuracy and robustness are insufficient, and real-time automated processing is difficult to achieve.

[0007] With the rapid development of deep learning technology in image segmentation and object detection, introducing deep learning into the rapid detection and recognition of droplet images is expected to break through the performance bottleneck of traditional algorithms. However, existing deep learning droplet detection methods are mostly single-view two-dimensional image processing and have not yet been combined with multi-view three-dimensional reconstruction and single-frame multi-pulse temporal imaging. Summary of the Invention

[0008] In view of the above-mentioned technical problems, the purpose of this invention is to provide a method and system for detecting parameters of microdroplet jetting droplets in a multi-view single-frame stroboscopic manner. This method organically combines single-frame multi-pulse temporal imaging, multi-view three-dimensional reconstruction, multi-view consistency constraints, and deep learning image processing technology. It can simultaneously acquire the three-dimensional spatiotemporal information (three-dimensional shape, volume, and velocity) of the main droplet during microdroplet jetting additive manufacturing in a single exposure. Based on this, it can verify the reliable synchronous detection and identification of satellite droplets by realizing three-dimensional ellipsoidal shape constraints and independent trajectory continuity.

[0009] To achieve the above-mentioned objectives, a multi-view, single-frame stroboscopic microdroplet ejection parameter detection system is provided. The system utilizes microdroplet ejection additive manufacturing technologies, including at least piezoelectric inkjet additive manufacturing and electrohydrodynamic inkjet additive manufacturing technologies. The system comprises: Droplet ejection module, used for ejecting droplets; The multi-view image acquisition module is used to synchronously and controllably acquire single-frame multi-pulse images of the ejected droplets from multiple perspectives using single-frame multi-pulse stroboscopic illumination technology. The droplet images at different pulse moments are all laterally displaced in the single-frame multi-pulse images of each perspective according to the same time sequence. The parameter detection module preprocesses all single-frame multi-pulse images, then combines subpixel localization and semantic segmentation to identify droplet images. After verifying the droplet images using multi-view temporal consistency based on lateral displacement, it employs a combination of epipolar geometric constraints and deep learning to perform cross-view feature matching on the verified reliable droplet images. Based on the matching results, it performs 3D reconstruction of the droplets. A lightweight classification network is used to filter and distinguish between the main droplet and candidate satellite droplets in the image. Parameters are calculated based on the 3D reconstruction results of the main droplet. For candidate satellite droplets, 3D ellipsoidal morphology constraints and independent trajectory continuity verification are performed to determine the real satellite droplets. Finally, parameters are calculated based on the 3D reconstruction results of the real satellite droplets.

[0010] Preferably, multiple image acquisition units are used to simultaneously acquire single-frame multi-pulse images of the ejected droplets from various perspectives using single-frame multi-pulse stroboscopic illumination technology; Each image acquisition unit includes an industrial camera, an imaging lens, a programmable stroboscopic illumination source, and a piezoelectric deflector. All industrial cameras are synchronously triggered to start exposure and generate a pulse sequence synchronized with the exposure window. This drives each programmable stroboscopic illumination source to generate multiple short stroboscopic pulses during a single exposure. All imaging lenses synchronously image the ejected droplets. Each piezoelectric deflector is synchronously triggered and uniformly controlled, deflecting synchronously with the short stroboscopic pulses. This ensures that the droplet images at different pulse moments undergo lateral displacement in the same time sequence on the single-frame multi-pulse image from each viewing angle.

[0011] Preferably, the industrial camera has a maximum frame rate of no more than 200 fps, an adjustable exposure time range of 1 μs to 1s, and external trigger and global shutter functions. The imaging lens is a telecentric lens or a continuously zoom microscope lens, and the working distance is adjustable. The single pulse width of the programmable strobe lighting source is no greater than 10 μs, the interval between adjacent pulses is programmable and can be set from 5 μs to 2000 μs, and the number of pulses... ≥2; All image acquisition units are evenly distributed in a circumferential direction perpendicular to the direction of droplet descent, and use the same pitch angle.

[0012] Preferably, the method combines sub-pixel localization and semantic segmentation to recognize droplet images, including: The sub-pixel gray-level centroid coordinates of each droplet image in the droplet image region are calculated using the gray-level centroid method. Then, the sub-pixel precision contour of each droplet image is extracted using the Canny edge detection operator combined with the Zernike moment sub-pixel edge localization algorithm. A deep learning-based pre-trained semantic segmentation network is used to perform end-to-end droplet pixel-level detection and segmentation on single-frame multi-pulse images from various perspectives. The droplet image is determined by combining the subpixel precision contour of the droplet image with the pixel-level detection and segmentation results of the droplet.

[0013] Preferably, the droplet image is subjected to multi-view temporal consistency verification based on lateral displacement, including: For a sequence of droplet images arranged in chronological order within the same viewpoint, the displacement vector of the droplet images at adjacent time points based on lateral displacement is calculated. By detecting abrupt changes or anomalies in the displacement vector, it is preliminarily determined whether there is interference from satellite droplets or image segmentation error, and reliable droplet images without interference or error are identified.

[0014] Preferably, a combination of epipolar geometric constraints and deep learning is used to perform cross-view feature matching on the verified reliable droplet image, including: For the same pulse moment, the droplet image centers extracted from each viewpoint are matched across viewpoints using epipolar geometric constraints to obtain candidate matches; For droplet images at the same pulse moment, the verified droplet images detected from each viewpoint are input into a feature extraction network to obtain the depth feature vector of each droplet image. Subsequently, the cosine similarity of the depth feature vectors of droplet images between different viewpoints is calculated as a visual similarity score. At the same time, the epipolar distance between each candidate matching pair is calculated based on the basic matrix obtained from the bi-objective calibration as a geometric consistency score. The above two scores are weighted and fused to construct a matching cost matrix, where the lower the value of the matrix element, the smaller the comprehensive matching cost of the candidate match. The Hungarian algorithm is used to solve the matching cost matrix to find the optimal match. Under the premise of satisfying the "one-to-one" matching constraint, the global matching cost is minimized, and the optimal cross-view droplet image correspondence is obtained.

[0015] Preferably, the candidate satellite droplets are subjected to three-dimensional ellipsoidal morphology constraints and independent trajectory continuity verification to determine the real satellite droplets, including: First, morphological constraints are applied to the 3D reconstruction results of each candidate satellite droplet: if the semi-axis ratio of the 3D reconstruction of a candidate satellite droplet exceeds a preset range, or its volume is smaller than a preset ratio threshold of the main droplet volume, it is further screened as a secondary candidate satellite droplet; for secondary candidate satellite droplets, it is further verified whether their 3D coordinates at multiple consecutive pulse moments constitute an independent continuous trajectory. If their trajectory is spatially distinguishable from the main droplet trajectory and satisfies motion continuity, they are confirmed as real satellite droplets.

[0016] Preferably, when calculating the three-dimensional velocity parameters, a time-series trajectory smoothing reconstruction based on the spatiotemporal continuity of flight is adopted: relying on the fact that the flight trajectory of the droplet in a short time satisfies the second-order kinematic model, the three-dimensional coordinate sequence of all pulse moments is fitted as a whole into a three-dimensional spatial trajectory, the overall least squares fitting method is introduced, and the sum of squares of the vertical distances from all three-dimensional coordinates to the fitted trajectory is minimized to obtain the optimal trajectory estimate; based on the fitted optimal trajectory estimate, the theoretical coordinate values ​​of each pulse moment are recalculated and used to replace the original observed three-dimensional coordinates for velocity calculation.

[0017] Preferably, preprocessing is performed on all single-frame multi-pulse images, including: For each viewpoint, the single-frame multi-pulse image is sequentially processed by denoising, adaptive contrast enhancement, threshold segmentation, and connected component analysis to extract the image regions of each droplet.

[0018] To achieve the above-mentioned objectives, a method for detecting microdroplet ejection parameters using multi-view single-frame stroboscopic imaging is also provided, comprising the following steps: Droplets are ejected using a droplet ejection module; The multi-view image acquisition module is used to acquire single-frame multi-pulse images from various viewpoints, and the droplet images at different pulse moments all exhibit lateral displacement in the single-frame multi-pulse images from various viewpoints according to the same time sequence. After preprocessing all single-frame multi-pulse images using a parameter detection module, droplet images are identified by combining subpixel localization and semantic segmentation. Multi-view temporal consistency verification based on lateral displacement is performed on the droplet images. Then, a combination of epipolar geometric constraints and deep learning is used to perform cross-view feature matching on the verified reliable droplet images. Based on the matching results, 3D reconstruction of the droplets is performed. A lightweight classification network is used to screen and distinguish between the main droplet and candidate satellite droplets in the image. Parameter calculation is performed based on the 3D reconstruction results of the main droplet. For the candidate satellite droplets, 3D ellipsoidal morphology constraints and independent trajectory continuity verification are performed to determine the real satellite droplets. Finally, parameter calculation is performed based on the 3D reconstruction results of the real satellite droplets.

[0019] Compared with the prior art, the beneficial effects of the present invention include at least the following: (1) Low-cost high-speed 3D detection: This invention uses a common industrial camera with a frame rate of no more than 200 fps and achieves sub-millisecond temporal resolution (interval between adjacent pulses) in a single exposure through single-frame multi-pulse stroboscopic technology. The speed can be as small as tens of microseconds, enabling the measurement of high-speed droplet velocity without the need for expensive high-speed cameras. Furthermore, by employing an orthogonal or equiangular arrangement of at least two viewpoints, the true three-dimensional information of the droplet is obtained through multi-view fusion, overcoming the fundamental limitation of single-view detection, which can only acquire two-dimensional projection information.

[0020] (2) Spatiotemporal integrated detection and synchronous measurement with multi-view and single-frame multi-pulse collaboration: This invention embeds single-frame multi-pulse stroboscopic technology into a multi-view imaging architecture, enabling each single-frame image from each viewpoint to simultaneously carry information in both the temporal and spatial dimensions. Through cross-view temporal matching and three-dimensional geometric reconstruction, the complete motion trajectory of the droplet in three-dimensional space can be obtained within a single exposure, simultaneously achieving high-precision measurement of droplet volume and three-dimensional velocity. This information acquisition method of "single exposure = three-dimensional space + time series" eliminates the need for multiple acquisitions or switching of working modes, resulting in high detection efficiency and good data consistency. This is something that cannot be achieved by simple multi-view multi-frame acquisition or simple single-view single-frame multi-pulse schemes, demonstrating significant technological advancement.

[0021] (3) Reliable identification and quantitative detection of satellite droplets based on multi-view temporal matching and morphological constraints: This invention utilizes synchronous imaging at the same pulse moment from at least two perspectives. Through multi-view information fusion, connected component analysis, epipolar constraints, and three-dimensional spatial consistency verification, it can spatially separate the main droplet and satellite droplets, accurately identifying the presence of satellite droplets. Simultaneously, it introduces three-dimensional ellipsoidal morphological constraints and independent trajectory continuity verification to double-confirm candidate satellite droplets and quantitatively measure the volume, velocity, and spatial distribution of each satellite droplet. Compared with the existing single-view scheme that judges satellite droplets based solely on two-dimensional morphological thresholds, the multi-view temporal consistency constraints and morphological-trajectory dual constraints of this invention can effectively eliminate misjudgments caused by factors such as projection overlap and image noise, significantly improving the accuracy of satellite droplet identification and overcoming the limitation of ineffective identification when the projections of satellite droplets and main droplets overlap in single-view detection.

[0022] (4) Deep learning enables rapid detection and recognition: This invention introduces deep learning semantic segmentation network and lightweight classification network into the multi-view single-frame multi-pulse imaging architecture, realizing end-to-end rapid detection of droplet images, cross-view robust matching and intelligent initial screening of satellite droplets. It overcomes the problem of insufficient segmentation accuracy of traditional algorithms in complex scenarios such as droplet overlap and blurred boundaries, and takes into account both detection speed and accuracy, enabling the system to have real-time online processing capabilities.

[0023] (5) High measurement accuracy and strong robustness: The present invention adopts sub-pixel edge extraction and weighted ellipse fitting algorithm to improve the positioning accuracy of droplet contour to the sub-pixel level, and suppresses the interference of edge noise and incomplete contour areas through curvature adaptive weight allocation. Even when the droplet image has a certain degree of blur or insufficient contrast, it can still maintain stable measurement accuracy.

[0024] (6) The system has a simple structure and good compatibility: The multi-view image acquisition part is based on a general industrial camera and a conventional lens. There are no special customized parts, and it is easy to integrate and upgrade on the existing micro-droplet jetting platform. Upgrading from a single view to a dual view only requires adding a camera-lens-light source unit and synchronous control expansion. The changes are small and the engineering implementation cost is controllable.

[0025] (7) Strong real-time performance: The present invention completes the spatiotemporal information acquisition with a single frame image, without the need for multi-frame stitching, and the image processing flow is simplified, which is suitable for online real-time detection needs. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a schematic diagram of the structure of the multi-view single-frame stroboscopic microdroplet ejection parameter detection system provided in the embodiment; Figure 2 This is a schematic diagram of the structure of the multi-view image acquisition module provided in the embodiment; Figure 3 This is a flowchart of the multi-view single-frame stroboscopic microdroplet ejection parameter detection method provided in the embodiment; Figure 4 This is a comparison diagram of observed droplet morphology provided in the embodiment (arrows indicate satellite droplets). Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of this invention.

[0029] This invention provides a method and system for detecting parameters of microdroplet ejection using multi-view single-frame stroboscopic illumination. It employs at least two orthogonally or equiangularly arranged image acquisition units, each using a standard industrial camera with a frame rate not exceeding 200 fps. Combined with single-frame multi-pulse stroboscopic illumination technology, each camera records multiple images of the same droplet at different times within a single exposure. This allows for the simultaneous acquisition of spatiotemporal information from multiple perspectives in a single acquisition, achieving high-precision, low-cost, and synchronous detection of droplet 3D morphology reconstruction, volume, 3D velocity, and satellite droplets. In this process, the piezoelectric deflector in each image acquisition unit deflects synchronously with the stroboscopic pulse, so that the droplet images at different pulse moments are all laterally offset in the images from each viewpoint according to the same time sequence, which facilitates cross-viewpoint temporal matching. In the image processing stage, a deep learning semantic segmentation network is introduced to perform end-to-end droplet image detection and segmentation on single-frame multi-pulse images from each viewpoint, and a deep learning feature matching network is combined with epipolar constraints to complete robust cross-viewpoint matching. Through multi-view geometric reconstruction and three-dimensional ellipsoidal morphological constraints, combined with trajectory smoothing reconstruction based on the spatiotemporal continuity of flight, accurate estimation of the three-dimensional motion trajectory of the droplet and confirmation of the multi-view consistency of the satellite droplet are achieved.

[0030] like Figure 1 As shown, the embodiment provides a multi-view single-frame stroboscopic microdroplet ejection parameter detection system 100, including a droplet ejection module 110, a multi-view image acquisition module 120, and a parameter detection module 130. The droplet ejection module 110 is used to eject droplets and specifically includes a piezoelectric nozzle and a drive controller. The piezoelectric nozzle is controlled by the drive controller to eject droplets according to a preset drive waveform.

[0031] In this embodiment, the multi-view image acquisition module 120 is used to synchronously and controllably acquire single-frame multi-pulse images of the ejected droplet from multiple perspectives using single-frame multi-pulse stroboscopic illumination technology. Specifically, at least two image acquisition units are used to acquire single-frame multi-pulse images of each perspective. All image acquisition units are evenly distributed in a circumferential direction perpendicular to the droplet's descent direction and use the same pitch angle, thus ensuring that the optical axes of all image acquisition units intersect in the central region of the droplet's flight path. When two image acquisition units are used, the two optical axes are arranged orthogonally (i.e., the included angle is 90°), and when three image acquisition units are used, the included angle between adjacent optical axes is 120°.

[0032] Each image acquisition unit includes an industrial camera, an imaging lens, a programmable stroboscopic illumination source, and a piezoelectric deflector. All industrial cameras are synchronously triggered to begin exposure, generating a pulse sequence synchronized with the exposure window. This drives each programmable stroboscopic illumination source to generate multiple short-duration stroboscopic pulses during a single exposure. The time alignment error of each pulse sequence is less than 20 ns to ensure temporal synchronization of images from different viewpoints. All imaging lenses synchronously image the ejected droplets. Each piezoelectric deflector is uniformly controlled by synchronous triggering and operates synchronously with the short-duration stroboscopic pulses. The piezoelectric deflectors at all viewpoints deflect simultaneously according to the same time sequence, deflecting by a small angle with each pulse. This ensures that the droplet images at different pulse moments exhibit lateral displacement according to the same time sequence in the single-frame multi-pulse images from each viewpoint. This multi-viewpoint collaborative deflection strategy avoids overlap of multiple droplet images within a single viewpoint while ensuring consistent lateral offset of droplet images at the same time across different viewpoints. This facilitates spatial correspondence for subsequent cross-viewpoint temporal feature matching based on lateral offset and is beneficial for 3D reconstruction.

[0033] In this embodiment, the industrial camera is a global shutter CCD or CMOS camera with a maximum frame rate of no more than 200 fps and an adjustable exposure time range of 1 μs to 1 s. It features external triggering and a global shutter function. The global shutter characteristic ensures that all pixels are exposed simultaneously during stroboscopic illumination, avoiding distortion caused by rolling shutters. The imaging lens is a telecentric lens or a continuous zoom microscope lens with an adjustable working distance, adaptable to the observation of droplets of different sizes. The programmable stroboscopic illumination source is a high-brightness LED driven by a programmable pulse generator, capable of generating short light pulses with a single pulse width ≤10 μs. The programmable setting range for the interval between adjacent pulses is 5 μs to 2000 μs, and the number of pulses is [not specified]. ≥2. Short pulses can effectively freeze high-speed flying microdroplets, eliminating motion blur.

[0034] In this embodiment, the parameter detection module 130 preprocesses all single-frame multi-pulse images, combines subpixel localization methods and semantic segmentation to identify droplet images, performs multi-view temporal consistency verification on the droplet images based on lateral displacement, performs cross-view feature matching on the verified reliable droplet images using a combination of epipolar geometric constraints and deep learning, performs 3D reconstruction of the droplets based on the matching results, uses a lightweight classification network to screen and distinguish between the main droplet and candidate satellite droplets in the image, calculates parameters based on the 3D reconstruction results of the main droplet, performs 3D ellipsoidal morphology constraints and independent trajectory continuity verification on the candidate satellite droplets to determine the real satellite droplets, and calculates parameters based on the 3D reconstruction results of the real satellite droplets.

[0035] In this embodiment, all single-frame multi-pulse images are preprocessed, including: performing denoising, adaptive contrast enhancement, threshold segmentation, and connected component analysis sequentially on the single-frame multi-pulse images from each viewpoint to extract the image regions of each droplet. Median filtering is preferably used for denoising. Adaptive contrast enhancement preferably uses adaptive histogram equalization (CLAHE) to improve the contrast between the droplets and the background.

[0036] For threshold segmentation, a hybrid strategy combining Otsu's global thresholding method and local adaptive thresholding method is adopted: First, the global optimal threshold in the contrast-enhanced image is calculated using Otsu's global thresholding method for initial segmentation. For regions with low global segmentation confidence (uncertain foreground regions with grayscale values ​​close to the threshold), local adaptive thresholding is used for fine-tuning to separate the segments. A binary image.

[0037] For connected component analysis, the binarized image after threshold segmentation is labeled with connected components, dividing interconnected droplet image pixel regions into independent connected components, each corresponding to a droplet image candidate region. Connected components with areas too small (less than a preset pixel threshold) are identified as noise regions and removed; connected components with areas too large are identified as multiple droplet images adhering together and separated using a watershed algorithm or concave point segmentation method. Through connected component analysis, the following is obtained: Each droplet image region serves as the basic unit for subsequent subpixel localization and contour extraction.

[0038] In this embodiment, a subpixel localization method and semantic segmentation are combined to identify droplet images. Specifically, the subpixel localization method includes: first, calculating the subpixel gray-level centroid coordinates of each droplet image in the droplet image region using the gray-level centroid method; then, extracting the subpixel precision contour of each droplet image using the Canny edge detection operator combined with the Zernike moment subpixel edge localization algorithm. Specifically, this includes: extracting pixel-level edges using the Canny operator, calculating the Zernike moment for each edge point in its neighborhood, obtaining background gray-level and edge height information, and inferring the subpixel coordinate position of the edge through the analytical relationship between geometric moments and edge parameters to obtain the subpixel precision droplet contour. Finally, the droplet contour is fitted: ellipse fitting is performed on the extracted subpixel edge point set using the weighted least squares ellipse fitting method. The weighted least squares objective function is... for: in, The number of sub-pixel edge points, weight The curvature characteristics and local contrast of the edge points determine the grouping of edge points into "main arc segments" and "end arc segments" based on their curvature values. Points in the main arc segments are assigned higher weights, while those in the end arc segments are assigned lower weights. By solving the aforementioned weighted least squares problem, the ellipse fitting parameters, including the ellipse center and major axis length, are obtained. minor axis length Direction angle . These are the parameters of the general equation of an ellipse.

[0039] A deep learning-based pre-trained semantic segmentation network is employed for end-to-end pixel-level droplet detection and segmentation of single-frame multi-pulse images from various perspectives. Specifically, the pre-trained semantic segmentation network is input into single-frame multi-pulse images from various perspectives, and outputs a pixel-level mask for each droplet image to achieve rapid detection and coarse localization of the droplet image region. During the training phase, the semantic segmentation network uses a droplet image dataset acquired simultaneously from multiple perspectives. The annotation information includes the contour mask and time sequence number of each droplet image, where the time sequence number is a unique identifier that distinguishes droplet images at different times in the same image during training. Through the rapid inference of the semantic segmentation network, the detection of all droplet images can be completed within milliseconds, significantly improving the processing speed and adaptability to complex scenes (droplet overlap, blurred boundaries, uneven illumination). The semantic segmentation results and sub-pixel localization methods complement each other. The semantic segmentation network provides fast pixel-level coarse localization, while traditional sub-pixel methods provide high-precision edge localization and contour fitting. The fusion of the two yields droplet images that balance speed and accuracy.

[0040] In this embodiment, to avoid misjudgments caused by projection overlap in single-view detection, multi-view temporal consistency verification of the droplet images is also required. For a sequence of droplet images arranged chronologically within the same viewpoint, the displacement vectors of droplet images at adjacent moments based on lateral displacement are calculated. Since the motion of droplets within the same viewpoint has spatiotemporal continuity, the displacement vectors at each moment should satisfy smoothness constraints. By detecting abrupt changes or anomalies in the displacement vectors, it is possible to preliminarily determine whether there is interference from satellite droplets or image segmentation errors, and to identify reliable droplet images that are free from interference and errors, thereby improving the reliability of subsequent cross-view matching.

[0041] In this embodiment, after multi-view temporal consistency verification, a combination of epipolar geometric constraints and deep learning is used to perform cross-view feature matching on the verified reliable droplet image. Specifically, for the same pulse moment... The centers of the droplet images extracted from each viewpoint are matched using epipolar geometric constraints across viewpoints. Taking a dual-viewpoint model as an example, spatial points... Projection matrices on the left and right image planes and Satisfying the polar geometric constraint equations: in The basic matrix is ​​used. Triangulating the matched pixel pairs yields the three-dimensional spatial coordinates of the droplet at that moment. The specific calculation method for triangulation is as follows: Let the projection matrices of the left and right cameras be respectively... and (All are 3×4 matrices), the homogeneous coordinates of the matched pixels are: and Then the three-dimensional spatial coordinates The following system of linear equations can be obtained by solving them using methods such as the least squares method: in , , They represent The row vectors of rows 1, 2, and 3. , , They represent The row vectors of rows 1, 2, and 3.

[0042] When performing cross-view feature matching on verified droplet images using a deep learning-based approach, specifically, for droplet images at the same pulse moment, the verified droplet images detected from each viewpoint are input into a feature extraction network to obtain the depth feature vector of each droplet image. Subsequently, the cosine similarity of the depth feature vectors of droplet images from different viewpoints is calculated as a visual similarity score. Simultaneously, the epipolar distance between each candidate matching pair is calculated based on the fundamental matrix obtained from bi-objective calibration as a geometric consistency score. These two scores are then weighted and fused to construct a matching cost matrix, where lower element values ​​indicate a smaller overall matching cost for the candidate match. Based on this, the Hungarian algorithm is used to solve for the optimal match on the matching cost matrix, minimizing the global matching cost while satisfying the "one-to-one" matching constraint, thus obtaining the final cross-view droplet image correspondence. This invention employs a cross-view feature matching method combining epipolar geometric constraints and deep learning, which can provide additional discriminative information when droplet images are relatively similar or when there are multiple ambiguities in the epipolar geometric constraints, thereby improving the robustness and accuracy of cross-view matching.

[0043] In this embodiment, three-dimensional reconstruction of the droplet is performed based on the matching results. Specifically, ellipse parameters are obtained by ellipse fitting of the droplet images at the same pulse moment from various viewpoints, including: the length of the major axis. minor axis length Direction angle A three-dimensional ellipsoidal model of the droplet is reconstructed based on the multi-view geometric relationships and the elliptical fitting results of the matched droplet image. Specifically, taking a two-view example: Let the lateral magnification obtained from the system calibration be... (Unit: pixels / actual length), then when the droplet's orientation is perpendicular (along the major axis)... Under the approximation of the axis, the lengths of the three semi-axes of the three-dimensional ellipsoid , , It can be calculated using the following formula: , , Among them, the elliptic parameters of the droplet images from two different perspectives and When the droplet orientation angle is not zero, a pre-calibrated three-dimensional ellipsoid reconstruction model can be used. By substituting the elliptical parameters from the two perspectives, the lengths of the three semi-axes can be obtained through nonlinear least squares fitting. , , As a result of three-dimensional reconstruction representing three-dimensional morphology.

[0044] In this embodiment, a lightweight classification network is also used to filter and distinguish between the main droplet and candidate satellite droplets in the image. The droplet image region obtained from connected component analysis is input into the pre-trained lightweight classification network, which outputs a probability score indicating whether the droplet image belongs to the main droplet or a satellite droplet. Droplets with probability scores higher than a preset threshold are selected as candidate satellite droplets. This lightweight classification network can quickly filter out a large number of noise regions (such as dust, bubbles, and reflective points) that clearly do not belong to satellite droplets after connected component analysis, reducing the computational burden of subsequent 3D reconstruction and improving overall processing efficiency.

[0045] For the master droplet, parameters such as volume and velocity are calculated based on the 3D reconstruction results, specifically including: First, perform temporal correlation: based on the spatial order of the droplet images in the image (the jet direction is usually vertically downward, corresponding sequentially from high to low according to the vertical coordinates). , … Determine the temporal order of multiple droplet images within the same viewpoint, and determine the time label corresponding to each droplet image. .

[0046] Then, the three-dimensional velocity is calculated: the three-dimensional velocity components and the composite velocity are calculated based on the three-dimensional coordinate differences between adjacent pulse moments and the pulse time interval. To improve accuracy, the coordinate-time relationship of all pulse moments is fitted using the linear least squares method to obtain the average velocity. in, As the reference coordinates, Let i be the three-dimensional coordinates corresponding to the i-th time step. These are the velocity components in the x, y, and z directions, respectively.

[0047] In this embodiment, to effectively suppress single-point position measurement deviations caused by light intensity fluctuations, instantaneous changes in droplet attitude, or slight defocusing in a single frame image, and to improve the robustness and repeatability of velocity measurement, a temporal trajectory smoothing reconstruction based on the spatiotemporal continuity of flight is adopted. Relying on the fact that the droplet's flight trajectory in a short time satisfies a second-order kinematic model, the three-dimensional coordinate sequence of all pulse moments is obtained. The entire trajectory is fitted into a three-dimensional spatial trajectory. A global least squares fitting method is introduced, which minimizes the sum of squared perpendicular distances from all three-dimensional coordinates to the fitted trajectory to obtain the optimal trajectory estimate. Based on the fitted optimal trajectory estimate, the theoretical coordinate values ​​at each pulse moment are recalculated and used to replace the original observed three-dimensional coordinates for velocity calculation.

[0048] Finally, the volume is calculated: the droplet volume is calculated based on the semi-axis of the 3D ellipsoid in the 3D reconstruction results. For non-ellipsoidal droplets, such as trailing droplets or jet fracture transition morphologies, a multi-view profile fusion method is used to reconstruct the 3D surface mesh, and then the volume is calculated by integrating the mesh. For standard spherical or near-spherical droplets, the equivalent diameter method can be directly used for verification.

[0049] In this embodiment, the candidate satellite droplets are further subjected to three-dimensional ellipsoidal morphological constraints and independent trajectory continuity verification to determine the real satellite droplets. Specifically, this includes: firstly, applying morphological constraints to the three-dimensional reconstruction results of each candidate satellite droplet: the main droplet should satisfy the ellipsoidal morphological constraints in three-dimensional space, and its three semi-axis... The relative proportions of the two droplets should be within a preset range. If the semi-axis ratio of the candidate satellite droplet after 3D reconstruction exceeds the preset range, or its volume is smaller than the preset ratio threshold of the main droplet volume, it is further screened as a secondary candidate satellite droplet. For secondary candidate satellite droplets, it is further verified whether their 3D coordinates at multiple consecutive pulse moments constitute an independent continuous trajectory. If its trajectory is spatially distinguishable from the main droplet trajectory and satisfies motion continuity, it is confirmed as a real satellite droplet; otherwise, it is judged as noise or segmentation error and discarded.

[0050] The parameters of the three-dimensional reconstruction results of the real satellite droplets are calculated. Specifically, the volume and velocity of the real satellite droplets are calculated according to the calculation method of the master droplet.

[0051] In this embodiment, the final output includes the three-dimensional velocity vector and volume of the main droplet, as well as the semi-axis and roundness parameters of the three-dimensional ellipsoid representing its three-dimensional morphology; the number of satellite droplets, the volume, velocity, and spatial distribution of each satellite droplet are also output. The system is calibrated offline using standard-sized microspheres to obtain pixel-to-actual-size conversion coefficients and ellipsoid model correction factors. The measurement results are then corrected, and the number, three-dimensional spatial distribution, size, and velocity of the satellite droplets are output.

[0052] The embodiment provides a method for detecting parameters of microdroplet jetting droplets using multi-view single-frame stroboscopic imaging. This method applies the above-mentioned system and performs system calibration and parameter preset before detection.

[0053] For system calibration: Intrinsic parameters are calibrated for each camera separately to obtain focal length, principal point coordinates, and distortion coefficients. Then, multi-camera calibration is performed to obtain the rotation matrix and translation vector between any two cameras, establishing a 3D coordinate mapping relationship within a common field of view. A checkerboard calibration board is used, with a grid spacing of 0.05–0.2 mm. After calibration, the reprojection error should be less than 0.2 pixels.

[0054] For parameter presets: based on the estimated droplet velocity To determine the required measurement accuracy, set the following parameters: camera exposure time. The setting range is 1 μs to 1 s (typical value 10 to 15 ms); the number of strobe pulses per camera. ( ≥2, recommended (balancing measurement accuracy and image clarity); adjacent pulse time interval The setting range is 5 μs to 2000 μs (typical value 20 μs to 100 μs, determined based on droplet velocity and number of pulses); single pulse width The set range is 0.1 μs to 10 μs (typical value 0.5 μs to 2 μs). It must meet the following requirements. ,and ( (This refers to the droplet ejection frequency) to avoid droplet overlap caused by different ejection events within the same camera's field of view. If a piezoelectric deflector is used, the deflection angle increment corresponding to each pulse is set. This makes the lateral spacing between adjacent droplets on the image larger than the droplet diameter.

[0055] After system calibration and parameter preset are completed, the parameters of the microdroplet ejection droplets are detected, such as... Figure 3 As shown, the detection method includes the following steps: S1, using the droplet ejection module to eject droplets; S2, using a multi-view image acquisition module to acquire single-frame multi-pulse images from various viewpoints, and the droplet images at different pulse moments all exhibit lateral displacement in the single-frame multi-pulse images from various viewpoints according to the same time sequence. S3, after preprocessing all single-frame multi-pulse images using the parameter detection module, combines sub-pixel localization and semantic segmentation to identify droplet images. After verifying the consistency of droplet images based on lateral displacement using multi-view temporal consistency, a combination of epipolar geometric constraints and deep learning is used to perform cross-view feature matching on the verified reliable droplet images. Based on the matching results, 3D reconstruction of the droplets is performed. A lightweight classification network is used to screen and distinguish between the main droplet and candidate satellite droplets in the image. Parameters are calculated based on the 3D reconstruction results of the main droplet. After verifying the candidate satellite droplets by 3D ellipsoidal morphology constraints and independent trajectory continuity, the parameters are calculated based on the 3D reconstruction results of the real satellite droplets.

[0056] Based on the above system and method, the embodiment also provides an example of a dual-view orthogonal acquisition scheme. In terms of hardware configuration, the droplet jetting module uses a circular tube piezoelectric nozzle with an inner diameter of 50 μm, a jetting frequency of 500 Hz, a trapezoidal pulse driving waveform, and jets nano-silver conductive ink with a viscosity of 8 cP.

[0057] The first image acquisition unit includes a global shutter industrial CCD camera (resolution 1280×960, maximum frame rate 40 fps, global shutter, adjustable exposure time range of 5 μs to 200 ms, with external trigger function), a continuously variable telecentric lens (magnification adjustable range of 0.7× to 4.5×, set to 3× in this embodiment, working distance 95 mm, field of view approximately 2.5 mm × 2.0 mm), and a high-brightness blue LED strobe light source (wavelength 465 nm, single pulse width 1 μs, programmable pulse interval). The optical path of the first image acquisition unit also includes a piezoelectric deflector, whose deflection angle is controlled by synchronous triggering and operates synchronously with the strobe pulses.

[0058] The second image acquisition unit is identical in specifications to the first image acquisition unit. Both units are positioned at the same horizontal level. The optical axis of the first camera is along the X-direction, and the optical axis of the second camera is along the Y-direction. These two optical axes are perpendicular to each other and intersect at the center of the droplet's flight path. The distance from the lens to the observation position is kept equal in both directions to ensure consistent image scale. The second image acquisition unit is equipped with the same piezoelectric deflector as the first unit. Both piezoelectric deflectors are driven by the same control signal and deflect synchronously according to the same time sequence, ensuring that the droplet image at different pulse moments in each viewpoint undergoes lateral displacement in the same direction, facilitating subsequent cross-viewpoint time-series matching.

[0059] Synchronous Triggering and Strobe Control Module: Utilizes an FPGA board to generate an external trigger signal for the camera (frequency 40 Hz, synchronously divided with the nozzle's 500 Hz spray frequency, triggering an exposure every 12 or 13 spray cycles) and two pulse sequences. The camera exposure time is set to 1.8 ms. Pulse sequence parameters: =60 μs, =3, =1 μs, total pulse time 181 μs <2ms, satisfying the requirement that a single exposure contains only one ejection event. The alignment error between the two pulse edges is less than 10 ns.

[0060] The parameter detection module uses an industrial control computer to run the image processing and parameter calculation algorithm of this invention, and outputs the three-dimensional velocity, volume, roundness of the main droplet and satellite droplet information through a display or communication interface.

[0061] For parameter calibration, a checkerboard calibration board was used for camera intrinsic parameter calibration. 15-20 images of the calibration board in different poses were acquired for each of the two cameras. Zhang Zhengyou's calibration method was used to calculate the intrinsic parameter matrix (focal length, principal point) and distortion coefficients for each camera, completing distortion correction. For dual-camera calibration, the checkerboard calibration board was placed within a common field of view, and images of the calibration board were acquired simultaneously by both cameras. The rotation matrix R and translation vector T between the two cameras were calculated, establishing the mapping relationship between the world coordinate system and the pixel coordinate systems of each camera. After calibration, the reprojection error was less than 0.2 pixels.

[0062] Regarding parameter detection, the following parameters are preset: based on the estimated droplet velocity (approximately 3–5 m / s), the camera exposure time is set. =1.8 ms, number of pulses =3, pulse interval =60 μs, single pulse width =1 μs. Verification: =3×60 μs+1 μs=181 μs<2 ms, which meets the condition. Set the external trigger frequency of the camera to 40 Hz, that is, collect data once every 25 spray cycles.

[0063] Then, following steps S1-S3 above, image acquisition and parameter detection were performed. The detection results output the main droplet's three-dimensional velocity as approximately 3.6 m / s, volume as approximately 42 pL, and sphericity as approximately 0.96; satellite droplet information was also output. After 30 repeated measurements, the velocity standard deviation was <0.12 m / s, the volume standard deviation was <1.3 pL, and so on. Figure 4 The satellite droplet labeling results are shown.

[0064] The parameters of this invention can be flexibly adjusted to adapt to different spraying conditions: (1) For droplets with faster speeds (estimated >8 m / s), the pulse interval can be adjusted. The pulse count was reduced to 25 μs, the number of pulses N was increased to 5, more image points were recorded in the same exposure time, and the velocity measurement range was extended to 0–20 m / s.

[0065] (2) For smaller droplets (such as fly-level droplets generated by electrofluid jet), a telecentric lens with a higher magnification should be selected, and the strobe pulse width should be shortened to within 0.5 μs to eliminate motion blur and ensure edge detection accuracy.

[0066] (3) For jetting scenarios with a large number of satellite droplets, the time interval between adjacent pulses can be appropriately increased. This increases the distance between the main droplet and the satellite droplets in the image, making it easier to separate and independently identify connected components.

[0067] The specific embodiments described above illustrate the technical solution and beneficial effects of the present invention in detail. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-view, single-frame stroboscopic microdroplet ejection parameter detection system, characterized in that, The microdroplet jetting additive manufacturing technology used in the system includes at least piezoelectric inkjet additive manufacturing technology and electrohydrodynamic inkjet additive manufacturing technology, and the system includes: Droplet ejection module, used for ejecting droplets; The multi-view image acquisition module is used to synchronously and controllably acquire single-frame multi-pulse images of the ejected droplets from multiple perspectives using single-frame multi-pulse stroboscopic illumination technology. The droplet images at different pulse moments are all laterally displaced in the single-frame multi-pulse images of each perspective according to the same time sequence. The parameter detection module preprocesses all single-frame multi-pulse images, then combines subpixel localization and semantic segmentation to identify droplet images. After verifying the droplet images using multi-view temporal consistency based on lateral displacement, it employs a combination of epipolar geometric constraints and deep learning to perform cross-view feature matching on the verified reliable droplet images. Based on the matching results, it performs 3D reconstruction of the droplets. A lightweight classification network is used to filter and distinguish between the main droplet and candidate satellite droplets in the image. Parameters are calculated based on the 3D reconstruction results of the main droplet. For candidate satellite droplets, 3D ellipsoidal morphology constraints and independent trajectory continuity verification are performed to determine the real satellite droplets. Finally, parameters are calculated based on the 3D reconstruction results of the real satellite droplets.

2. The multi-view single-frame stroboscopic microdroplet ejection parameter detection system according to claim 1, characterized in that, Multiple image acquisition units are used to simultaneously acquire single-frame multi-pulse images of the jet droplets from various perspectives using single-frame multi-pulse stroboscopic illumination technology; Each image acquisition unit includes an industrial camera, an imaging lens, a programmable stroboscopic illumination source, and a piezoelectric deflector. All industrial cameras are synchronously triggered to start exposure and generate a pulse sequence synchronized with the exposure window. This drives each programmable stroboscopic illumination source to generate multiple short stroboscopic pulses during a single exposure. All imaging lenses synchronously image the ejected droplets. Each piezoelectric deflector is synchronously triggered and uniformly controlled, deflecting synchronously with the short stroboscopic pulses. This ensures that the droplet images at different pulse moments undergo lateral displacement in the same time sequence on the single-frame multi-pulse image from each viewing angle.

3. The multi-view single-frame stroboscopic microdroplet ejection parameter detection system according to claim 2, characterized in that, The industrial camera has a maximum frame rate of no more than 200 fps, an adjustable exposure time range of 1 μs to 1 s, and features external triggering and global shutter functions. The imaging lens is a telecentric lens or a continuously zoom microscope lens, and the working distance is adjustable. The single pulse width of the programmable strobe lighting source is no greater than 10 μs, the interval between adjacent pulses is programmable and can be set from 5 μs to 2000 μs, and the number of pulses... ≥2; All image acquisition units are evenly distributed in a circumferential direction perpendicular to the direction of droplet descent, and use the same pitch angle.

4. The multi-view single-frame stroboscopic microdroplet ejection parameter detection system according to claim 1, characterized in that, Combining subpixel localization methods and semantic segmentation to recognize droplet images includes: The sub-pixel gray-level centroid coordinates of each droplet image in the droplet image region are calculated using the gray-level centroid method. Then, the sub-pixel precision contour of each droplet image is extracted by combining the Canny edge detection operator with the Zernike moment sub-pixel edge localization algorithm. A deep learning-based pre-trained semantic segmentation network is used to perform end-to-end droplet pixel-level detection and segmentation on single-frame multi-pulse images from various perspectives. The droplet image is determined by combining the subpixel precision contour of the droplet image with the pixel-level detection and segmentation results of the droplet.

5. The multi-view single-frame stroboscopic microdroplet ejection parameter detection system according to claim 1, characterized in that, Multi-view temporal consistency verification of droplet images based on lateral displacement is performed, including: For a sequence of droplet images arranged in chronological order within the same viewpoint, the displacement vector of the droplet images at adjacent time points based on lateral displacement is calculated. By detecting abrupt changes or anomalies in the displacement vector, it is preliminarily determined whether there is interference from satellite droplets or image segmentation error, and reliable droplet images without interference or error are identified.

6. The multi-view single-frame stroboscopic microdroplet ejection parameter detection system according to claim 1, characterized in that, A combination of epipolar geometric constraints and deep learning was used to perform cross-view feature matching on validated reliable droplet images, including: For the same pulse moment, the droplet image centers extracted from each viewpoint are matched across viewpoints using epipolar geometric constraints to obtain candidate matches; For droplet images at the same pulse moment, the verified droplet images detected from each viewpoint are input into a feature extraction network to obtain the depth feature vector of each droplet image. Subsequently, the cosine similarity of the depth feature vectors of droplet images between different viewpoints is calculated as a visual similarity score. At the same time, the epipolar distance between each candidate matching pair is calculated based on the basic matrix obtained from the bi-objective calibration as a geometric consistency score. The above two scores are weighted and fused to construct a matching cost matrix, where the lower the value of the matrix element, the smaller the comprehensive matching cost of the candidate match. The Hungarian algorithm is used to solve the matching cost matrix for the optimal match, minimizing the global matching cost under the premise of satisfying the one-to-one matching constraint, and obtaining the optimal cross-view droplet image correspondence.

7. The multi-view single-frame stroboscopic microdroplet ejection parameter detection system according to claim 1, characterized in that, To determine the real satellite droplet, three-dimensional ellipsoidal morphology constraints and independent trajectory continuity verification were performed on candidate satellite droplets, including: First, morphological constraints are applied to the 3D reconstruction results of each candidate satellite droplet: if the semi-axis ratio of the 3D reconstruction of a candidate satellite droplet exceeds a preset range, or its volume is smaller than a preset ratio threshold of the main droplet volume, it is further screened as a secondary candidate satellite droplet; for secondary candidate satellite droplets, it is further verified whether their 3D coordinates at multiple consecutive pulse moments constitute an independent continuous trajectory. If their trajectory is spatially distinguishable from the main droplet trajectory and satisfies motion continuity, they are confirmed as real satellite droplets.

8. The multi-view single-frame stroboscopic microdroplet ejection parameter detection system according to claim 1, characterized in that, When calculating the three-dimensional velocity parameters, a time-series trajectory smoothing reconstruction based on the spatiotemporal continuity of flight is adopted: relying on the second-order kinematic model of the droplet's flight trajectory in a short time, the three-dimensional coordinate sequence of all pulse moments is fitted as a three-dimensional spatial trajectory. The overall least squares fitting method is introduced, and the sum of squared vertical distances from all three-dimensional coordinates to the fitted trajectory is minimized to obtain the optimal trajectory estimate. Based on the fitted optimal trajectory estimate, the theoretical coordinate values ​​of each pulse moment are recalculated and used to replace the original observed three-dimensional coordinates for velocity calculation.

9. The multi-view single-frame stroboscopic microdroplet ejection parameter detection system according to claim 1, characterized in that, Preprocessing is performed on all single-frame multipulse images, including: For each viewpoint, the single-frame multi-pulse image is sequentially processed by denoising, adaptive contrast enhancement, threshold segmentation, and connected component analysis to extract the image regions of each droplet.

10. A method for detecting parameters of microdroplet jetting liquid in a multi-view, single-frame stroboscopic pattern, characterized in that, The method utilizes microdroplet jetting additive manufacturing technologies, including at least piezoelectric inkjet additive manufacturing and electrohydrodynamic inkjet additive manufacturing technologies, and the method includes the following steps: Use a droplet ejection module to eject droplets; The multi-view image acquisition module is used to acquire single-frame multi-pulse images from various viewpoints, and the droplet images at different pulse moments all exhibit lateral displacement in the single-frame multi-pulse images from various viewpoints according to the same time sequence. After preprocessing all single-frame multi-pulse images using a parameter detection module, droplet images are identified by combining subpixel localization and semantic segmentation. Multi-view temporal consistency verification based on lateral displacement is performed on the droplet images. Then, a combination of epipolar geometric constraints and deep learning is used to perform cross-view feature matching on the verified reliable droplet images. Based on the matching results, 3D reconstruction of the droplets is performed. A lightweight classification network is used to screen and distinguish between the main droplet and candidate satellite droplets in the image. Parameter calculation is performed based on the 3D reconstruction results of the main droplet. For the candidate satellite droplets, 3D ellipsoidal morphology constraints and independent trajectory continuity verification are performed to determine the real satellite droplets. Finally, parameter calculation is performed based on the 3D reconstruction results of the real satellite droplets.

Images