Intelligent particle detection method and system based on laser confocal three-dimensional imaging

Abstract

This invention provides an intelligent particle detection method and system based on laser confocal three-dimensional imaging, applied in the field of particle detection technology. The method includes: filtering a liquid sample to be tested, trapping insoluble particles on the filter membrane surface and / or the filter membrane pore area; pre-scanning the filter membrane reference area and establishing a local three-dimensional reference surface; performing confocal three-dimensional data acquisition on the filter membrane surface; mapping continuous three-dimensional data to the corresponding local three-dimensional reference surface of the filter membrane, extracting independent protrusions; performing discrimination between real particles and interfering objects, identification of real particle categories, and verification of low-confidence objects based on three-dimensional combined features; and performing statistical aperture mapping of real particles according to preset detection standards and outputting the detection results. The system includes a central control unit, a sample preparation unit, a confocal imaging unit, and an image analysis unit. This invention can improve the accuracy and automation level of particle extraction, identification, classification, and statistical output against complex filter membrane backgrounds.

Description

Technical Field

[0001] This invention belongs to the field of particle detection technology, specifically relating to an intelligent particle detection method and system based on laser confocal three-dimensional imaging. Background Technology

[0002] Currently, the detection of insoluble microparticles in pharmaceuticals, injectable solutions, electronic-grade water, semiconductor chemical solutions, and other liquid samples typically employs techniques such as optical obscuration, light scattering, and microscopy. Among these, optical obscuration or light scattering is a commonly used automated detection method. Its basic principle involves allowing the liquid to flow through a detection area; when particles in the liquid pass through the area affected by the light beam, changes in the amount of light obstruction or the scattering signal are caused. The size and number of particles are then calculated based on these signal changes. This type of method offers fast detection speed and a high degree of automation, meeting the needs of large-scale testing of routine samples. Microscopy is often used as an arbitration or verification method. It typically involves filtering the liquid to trap insoluble microparticles on the filter membrane surface, then observing, identifying, and counting the particles on the filter membrane using an optical microscope to obtain the number and partial morphological information of the particles.

[0003] However, existing detection technologies still have certain limitations in practical applications. For colored, turbid, viscous, emulsified, or complex background liquid samples, optical obscuration and light scattering methods are easily affected by the optical properties of the sample itself, leading to increased fluctuations in the detection signal. This can result in misclassifying non-target objects such as bubbles, droplets, and matrix fluctuations as particles, or failing to adequately identify particles with low transparency and small size. Especially in complex sample systems, relying solely on changes in the light signal to indirectly infer particles is insufficient to accurately reflect the true state and spatial morphology of the particles, thus affecting the accuracy of the detection results.

[0004] On the other hand, while traditional microscopy can directly observe particles trapped on filter membranes and obtain surface information such as particle color and outline, its detection process typically involves multiple steps including filtration, drying, transfer, microscopic examination, and manual counting, resulting in low detection efficiency and a high dependence on operator experience. When the filter membrane surface has pore undulations, fiber textures, local warping, or residual liquid films, observers can easily confuse background structures, burrs, and adhering dust with actual particles, leading to poor reproducibility. Furthermore, ordinary optical microscopes primarily obtain two-dimensional image information, making it difficult to effectively characterize three-dimensional morphological parameters such as particle height, true surface area, volume, top curvature, and surface roughness. Therefore, it is insufficient to meet the needs for fine particle identification and source analysis under complex sample conditions.

[0005] Therefore, improving the accuracy, automation, and stability of insoluble particulate matter detection under complex background conditions has become a pressing technical problem in this field. Summary of the Invention

[0006] In view of the above-mentioned problems in the prior art, the purpose of this invention is to provide an intelligent particle detection method based on laser confocal three-dimensional imaging, which can improve the accuracy and automation level of particle extraction, identification, classification and statistical output under complex filter membrane background.

[0007] A smart particle detection method based on laser confocal 3D imaging includes the following steps: S1. Sample loading and filter membrane carrier preparation: Obtain the liquid sample to be tested, filter the liquid sample to be tested so that insoluble particles are trapped on the surface of the filter membrane and / or the pore area of ​​the filter membrane, and dry, transport and position the filter membrane to establish the initial spatial coordinates of the filter membrane relative to the equipment coordinate system. S2. Confocal 3D Data Acquisition of Filter Membrane Surface: A pre-scan is performed on the reference area on the filter membrane, and a local 3D reference surface of the filter membrane is established based on the 3D data of the reference area; on the basis of the local 3D reference surface of the filter membrane, a coarse scan is performed on the entire filter membrane to obtain the regional distribution results, and then an adaptive confocal fine scan is performed on different regions according to the regional distribution results to obtain a continuous 3D dataset of the filter membrane surface. S3. Intelligent particle analysis based on 3D morphology: The continuous 3D dataset is mapped to the corresponding local 3D reference surface of the filter membrane. Candidate foreground voxels are extracted based on the difference between the voxel height value and the reference surface height value. Independent protrusion objects are extracted through 3D connected component analysis, envelope morphology analysis, boundary closure analysis, and contact relationship analysis. 3D combined features are extracted from the independent protrusion objects, and the independent protrusion objects are distinguished as real particles and interference objects, and the real particle category is identified based on the 3D combined features. For independent protrusion objects with low confidence in the discrimination result, a local rescan is performed, and the discrimination is re-performed based on the rescan data. S4. Statistical Output of Detection Results: Based on the preset detection standards, the three-dimensional parameters of real particles are mapped to statistical caliber parameters, the real particles are classified and statistically analyzed, and the detection results are output.

[0008] Preferably, in step S1, a constant flow filtration method is used when filtering the liquid sample to be tested, and after filtration, a clean and dry airflow is introduced into the surface of the filter membrane to gently dry the filter membrane until the continuous liquid film on the surface of the filter membrane disappears or the thickness of the residual liquid film is lower than a preset threshold.

[0009] Preferably, the establishment of a local three-dimensional reference surface for the filter membrane includes: The reference area's 3D data is subjected to noise reduction, background correction, and intensity equalization. The reference area is divided into multiple local windows according to the preset window size; In each local window, a set of surface points on the exposed filter membrane surface is extracted, and a local reference surface for the corresponding local window is obtained by fitting the set of surface points. Based on the reference height value, local roughness reference value, and curvature variation range of each local reference surface, adjacent local reference surfaces are spliced ​​together to form a local three-dimensional reference surface for the filter membrane.

[0010] Preferably, adaptive confocal fine scanning includes: Based on the height variation range, brightness variation range and suspected protrusion distribution density obtained from the coarse scanning of the filter membrane, the scanning area is divided into flat area, gradually changing area, high undulation area and suspected object dense area. For flat areas, fine scanning is performed using the first magnification and the first Z-axis step size; for gradually changing areas, fine scanning is performed using the first magnification and the second Z-axis step size; for areas with high undulations, fine scanning is performed using the second magnification and the third Z-axis step size; and for areas with dense suspected objects, fine scanning is performed using the second magnification and the scanning range is reduced. Among them, the second Z-axis step size is smaller than the first Z-axis step size, and the third Z-axis step size is smaller than the second Z-axis step size.

[0011] Preferably, the independent protrusion object meets the following conditions: at least some voxel height differences are greater than the object height threshold, an independent protrusion envelope is formed, and there is an identifiable contact boundary between it and the local three-dimensional reference surface of the filter membrane; In addition, for all voxel height differences within the local roughness reference value range, or for local connected regions where there are local protrusions but the boundaries are continuously transitioned with the background and do not form an independent envelope, they are judged as filter membrane surface texture or filter membrane fiber undulation; for regions that are lower than the local reference surface as a whole, they are judged as filter membrane pores, depressions or membrane surface defects.

[0012] Preferably, the three-dimensional combined features include at least the maximum projected size, maximum height, object volume, true surface area, aspect ratio, sphericity, top curvature distribution, surface roughness, contact boundary closure with the local three-dimensional reference surface of the filter membrane, volume to projected area ratio, and grayscale variation continuity along the Z direction.

[0013] Preferably, the distinction between real particles and interference objects adopts a two-level discrimination method that combines a rule engine and a deep learning discrimination model; Among them, the rule engine performs initial screening of independent protrusion objects based on maximum height, contact boundary closure, surface roughness, aspect ratio, volume to projected area ratio, and the continuity of grayscale change along the Z direction. The deep learning discrimination model outputs discrimination results for real particles, bubble residues, dust adhesion, filter membrane burrs, or imaging artifacts based on the local three-dimensional voxel blocks, surface mesh data, and three-dimensional combined feature vectors corresponding to the independent protruding objects after initial screening.

[0014] Preferably, the local rescan includes: establishing a local scanning window based on the center coordinates and circumscribed dimensions of the low-confidence independent protrusion in the device coordinate system, switching to a high-magnification objective lens, and reducing the X-direction sampling interval, Y-direction sampling interval, and Z-direction step size to below the original fine scanning parameters; acquiring additional three-dimensional slice data of the top, sidewalls, and contact boundary area with the filter membrane around the low-confidence independent protrusion; and then re-judging based on the acquired data.

[0015] The second objective of this invention is to propose an intelligent particle detection system based on laser confocal three-dimensional imaging, comprising: The central control unit is used to establish detection tasks, generate control commands, and schedule the various functional units to work together. The sample preparation unit is used to acquire the liquid sample to be tested, filter the liquid sample to be tested so that insoluble particles are trapped on the surface of the filter membrane and / or the pore area of ​​the filter membrane, and dry, transport and position the filter membrane. The confocal imaging unit is used to pre-scan the reference area on the filter membrane. Based on the pre-scan results of the reference area, it helps to establish a local three-dimensional reference surface of the filter membrane. It also performs coarse scanning, adaptive fine scanning and local rescanning on the entire filter membrane to obtain continuous three-dimensional data of the filter membrane surface. The image analysis unit is used to extract independent protrusion objects from continuous three-dimensional data based on the local three-dimensional reference surface of the filter membrane, extract three-dimensional combined features from the independent protrusion objects, and perform discrimination between real particles and interference objects, identification of real particle categories, and statistical caliber mapping. The central control unit, sample preparation unit, confocal imaging unit, and image analysis unit are interconnected to execute an intelligent particle detection method based on laser confocal three-dimensional imaging.

[0016] Preferably, the confocal imaging unit includes a laser source, a confocal optical path system, an autofocus system, and a high-precision XYZ three-dimensional scanning stage; The image analysis unit is configured to: map continuous three-dimensional data to the corresponding local three-dimensional reference surface of the filter membrane, extract independent protrusion objects based on height difference calculation and three-dimensional connected component analysis; and perform real particle and interference object discrimination on the independent protrusion objects based on the maximum projected size, maximum height, object volume, real surface area, aspect ratio, sphericity, top curvature distribution, surface roughness, contact boundary closure, volume to projected area ratio, and grayscale change continuity along the Z direction.

[0017] The beneficial effects of this invention are: By filtering the liquid sample to be tested, insoluble particles are trapped on the surface of the filter membrane and / or in the pore area of ​​the filter membrane. Combined with filter membrane drying, transport and positioning processes, a particle carrier suitable for confocal imaging is constructed, which can improve the stability of the subsequent particle imaging process and provide a reliable basis for the automatic detection of insoluble particles.

[0018] By pre-scanning the filter membrane reference area and establishing a local three-dimensional reference surface of the filter membrane based on the three-dimensional data of the reference area, and then mapping the continuous three-dimensional dataset of the whole membrane to the corresponding local three-dimensional reference surface of the filter membrane, independent protrusion objects are extracted based on the voxel height difference, outer envelope morphology, boundary closure degree and contact relationship with the reference surface. This enables the effective differentiation between real particles and filter membrane surface texture, pore undulation and membrane surface defects, thereby improving the accuracy of particle extraction in complex filter membrane backgrounds.

[0019] By extracting three-dimensional combined features from independent protruding objects and performing discrimination between real particles and interference objects and identification of real particle categories based on the three-dimensional combined features, the ability to identify interference objects such as residual bubbles, filter membrane burrs, dust adhesion, and imaging artifacts can be improved, thereby enhancing the reliability of real particle detection results and improving the accuracy of particle classification analysis.

[0020] By classifying the scanning area according to the height variation range, brightness variation range, and suspected protrusion distribution density of different regions of the filter membrane, and applying different scanning parameters to different graded regions for adaptive confocal fine scanning, it is possible to reduce redundant scanning in flat areas while ensuring imaging accuracy in complex areas, thereby improving the efficiency of 3D data acquisition. Furthermore, performing local rescanning and re-discrimination on low-confidence independent protrusion objects can supplement the acquisition of objects with insufficient morphological information in the initial scan, improving the completeness and accuracy of the detection results.

[0021] By mapping the three-dimensional parameters of real particles to the statistical caliber parameters corresponding to preset detection standards, and performing categorized statistics and output of real particles, it is possible not only to obtain the quantity and particle size distribution results of insoluble particles, but also to output the category information and representative three-dimensional morphology information of real particles. This approach takes into account both the standard detection requirements and the needs of particle source analysis, thereby improving the practical value and traceability of the detection results. Attached Figure Description

[0022] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system block diagram of the present invention. Detailed Implementation

[0023] Example 1 like Figure 1 As shown, an intelligent particle detection method based on laser confocal three-dimensional imaging is applicable to the detection of insoluble particles in injection solutions, electronic-grade water, semiconductor chemical solutions, polishing slurries, and other liquid samples. The method specifically includes the following steps: S1. Sample loading and filter membrane carrier preparation S101. Place the liquid sample to be tested in the automatic sample injection position. The central control unit reads the sample number, sample name, detection volume, filter membrane specification, detection standard, and scanning parameters, and establishes a detection task corresponding to the liquid sample to be tested. The scanning parameters include at least the pre-scan magnification, fine scan magnification, pre-scan area size, local rescan area size, Z-axis scan start height, Z-axis scan end height, Z-axis step size range, independent protrusion object extraction threshold, true particle discrimination threshold, and low-confidence verification threshold.

[0024] The central control unit generates control commands based on the established detection task and sends them to the sample preparation unit, confocal imaging unit, and image analysis unit, respectively, so that each unit can complete the subsequent detection process under unified parameter conditions.

[0025] S102. The sample preparation unit draws the test liquid according to the detection volume set in the detection task and delivers the test liquid to the filtration device. The filtration device is equipped with a filter membrane for retaining insoluble particles. The test liquid passes through the filter membrane under driving pressure, the liquid components are discharged through the filter membrane, and insoluble particles are retained on the surface of the filter membrane and in the pore area of ​​the filter membrane. To reduce secondary migration of particles caused by liquid flow pulsation, it is preferable that the sample preparation unit performs filtration in a constant flow mode. After filtration, the sample preparation unit introduces a clean and dry airflow into the surface of the filter membrane to gently dry the filter membrane until the continuous liquid film on the filter membrane surface disappears or the thickness of the residual liquid film is lower than a preset threshold. After the above treatment, a particle carrier suitable for confocal imaging is formed on the surface of the filter membrane.

[0026] S103. The dried filter membrane is transported from the sample preparation unit to the object-carrying area of ​​the confocal imaging unit and placed on the high-precision XYZ three-dimensional scanning stage. Subsequently, the edges of the filter membrane are fixed by a clamping assembly to suppress positional drift during subsequent scanning. After the filter membrane is fixed, the central control unit controls the three-dimensional scanning stage to move to the preset origin position and establishes the initial spatial coordinates of the filter membrane relative to the device coordinate system through edge recognition, clamping reference recognition, or positioning mark recognition. This ensures that subsequent whole-membrane scanning, local re-scanning, and object verification are all performed in a unified spatial coordinate system.

[0027] S2, Confocal 3D Data Acquisition of Filter Membrane Surface S201. The central control unit first controls the confocal imaging unit to pre-scan the reference area on the filter membrane. The reference area is preferably a region with low particle density or an exposed filter membrane region. During the pre-scan, the autofocus system first determines the initial focal plane position of the reference area, and then the three-dimensional scanning stage drives the filter membrane to scan layer by layer along the Z direction at a preset reference step distance. At each Z position, the corresponding optical slice image of the reference area is acquired, and the three-dimensional data of the reference area is formed by combining the optical slice images corresponding to multiple Z positions.

[0028] After receiving the 3D data of the reference area, the image analysis unit sequentially performs noise reduction, background correction, and intensity equalization processing. Then, it divides the reference area into multiple local windows according to a preset window size. Within each local window, a set of surface points on the exposed filter membrane surface is extracted, and surface fitting is performed based on these points to obtain a local reference surface for that window. Subsequently, the reference height, local roughness reference value, and curvature variation range of each local reference surface are calculated. When the reference height is continuous between adjacent local windows and the curvature variation is within an allowable range, the local reference surfaces are stitched together to form a local 3D reference surface for the filter membrane.

[0029] Through the above processing, the entire filter membrane is no longer equivalent to a single plane. Instead, a background reference model with multiple local curved surfaces is constructed based on the actual undulation state of the filter membrane to adapt to the warping, pore undulation and surface texture changes that may occur after filtration and drying.

[0030] S202. After obtaining the local three-dimensional reference surface of the filter membrane, the central control unit controls the confocal imaging unit to perform a low-magnification coarse scan of the entire filter membrane. During the coarse scan, the three-dimensional scanning stage moves sub-region by sub-region according to the set path, and the confocal imaging unit acquires low-density three-dimensional slice data in each sub-region. The image analysis unit performs regional hierarchical processing on each sub-region based on the height variation range, brightness variation range, and suspected protrusion distribution density to form a regional distribution map.

[0031] The region distribution map includes at least flat areas, gradually changing areas, highly undulating areas, and areas with a high density of suspected objects. Preferably, when the height variation range of a region is below a first threshold and the density of suspected protrusions is below a second threshold, it is classified as a flat area; when the height variation range of a region is between the first and third thresholds, it is classified as a gradually changing area; when the height variation range of a region is above the third threshold, it is classified as a highly undulating area; and when the number of suspected protrusions in a region is above a fourth threshold, it is classified as an area with a high density of suspected objects. The image analysis unit sends the region distribution map to the central control unit for use during subsequent fine scanning.

[0032] S203. The central control unit uses the area distribution map obtained in step S202 to perform adaptive fine scanning by calling different scanning parameters for different level areas.

[0033] Specifically, for flat areas, the first set of scanning parameters is invoked, and a fine scan is performed using a first magnification and a first Z-axis step size; for gradually changing areas, the second set of scanning parameters is invoked, and a fine scan is performed using a first magnification and a second Z-axis step size, wherein the second Z-axis step size is smaller than the first Z-axis step size; for areas with high undulations, the third set of scanning parameters is invoked, and a fine scan is performed using a second magnification and a third Z-axis step size, wherein the third Z-axis step size is smaller than the second Z-axis step size; for areas with suspected dense objects, the scanning range is further reduced based on the corresponding area, and a fine scan is performed using a second magnification and a third Z-axis step size.

[0034] During the fine scanning process, the autofocus system first determines the starting focal plane for scanning the current sub-region, and then controls the 3D scanning stage to advance layer by layer along the Z direction. At each Z position, the laser beam performs XY plane scanning, and the photodetector collects the reflected light or fluorescence signal of the current focal plane to form an optical slice image corresponding to that Z position. Multiple optical slice images corresponding to Z positions constitute the fine scanning 3D data of the current sub-region. When a suspected object with a height exceeding a preset high threshold or a suspected object with a surface texture change exceeding a preset change threshold is detected in a certain sub-region, the central control unit marks the location of the suspected object and generates a list of objects to be rescanned for subsequent local rescanning.

[0035] S204. The image analysis unit performs 3D reconstruction and whole-film stitching on the fine-scan 3D data of each sub-region obtained in step S203. During reconstruction, the spatial coordinates in the X, Y, and Z directions are first recovered according to the acquisition position of each optical slice image. Then, the optical slice images of each layer are mapped to a unified voxel grid according to the preset voxel size to form 3D voxel data blocks of the corresponding sub-regions. During stitching, it is preferable to first perform coarse registration of adjacent sub-regions based on the displacement coordinates recorded by the 3D scanning stage, then perform height correction based on the height continuity of the local 3D reference surface of the filter membrane at the boundary of adjacent sub-regions, and finally perform lateral fine registration based on the texture similarity or grayscale matching degree of the boundary region, thereby stitching the 3D voxel data blocks of each sub-region into a continuous 3D dataset of the whole film.

[0036] When there are areas to be scanned in the list of areas to be scanned, after the local high-magnification scan is completed, the corresponding high-magnification scan data can be backfilled into the corresponding position of the whole membrane continuous three-dimensional dataset to form a locally enhanced whole membrane three-dimensional dataset.

[0037] S3. Intelligent particle analysis based on three-dimensional morphology S301, The image analysis unit extracts independent protrusion objects on the continuous three-dimensional dataset of the whole membrane, based on the local three-dimensional reference surface of the filter membrane established in step S201.

[0038] Specifically, the entire continuous 3D dataset is first mapped to the corresponding local 3D reference surface of the filter membrane using local windows. Then, height difference calculation is performed on the voxels within each local window to obtain the object height difference distribution, where the object height difference is the difference between the current voxel height value and the corresponding reference surface height value. Subsequently, voxels with height differences greater than a preset object height threshold are marked as candidate foreground voxels, and 3D connected component analysis is performed on the candidate foreground voxels to obtain multiple local connected components.

[0039] Furthermore, the outer envelope shape, boundary closure, and contact relationship with the reference surface are calculated for each locally connected region. When a locally connected region simultaneously satisfies the following conditions: at least some voxel height differences are greater than the object height threshold, an independent convex envelope is formed, and an identifiable contact boundary exists between it and the reference surface, the locally connected region is identified as an independent convex object. When all voxel height differences of a locally connected region are within the local roughness reference value range, or when a local convexity exists but its boundary transitions continuously with the background and does not form an independent envelope, it is judged as filter membrane surface texture or filter membrane fiber undulation, and is not recorded as an independent convex object. When a region is entirely lower than the local reference surface, it is judged as filter membrane pores, depressions, or membrane surface defects.

[0040] For objects partially embedded in the filter membrane pore area, their exposed height ratio and contact boundary closure are further calculated. When the exposed height ratio is greater than a preset ratio threshold and the contact boundary closure is greater than a preset closure threshold, they are still retained as independent protrusion objects.

[0041] S302, The image analysis unit extracts three-dimensional combined features from each independent protrusion object obtained in step S301. The three-dimensional combined features include at least the maximum projected size, equivalent projected diameter, maximum height, object volume, true surface area, aspect ratio, sphericity, top curvature distribution, surface roughness (Sa, Sq), contact boundary closure with the local three-dimensional reference surface of the filter membrane, volume to projected area ratio, and grayscale change continuity along the Z direction.

[0042] The maximum projected size is obtained by calculating the circumscribed dimension after projecting the independent protruding object along the Z-direction onto the XY plane; the maximum height is obtained by the height difference between the highest point of the independent protruding object and the corresponding local reference surface; the object volume is obtained by counting the number of object voxels and multiplying it by the volume of a single voxel; the true surface area is calculated after reconstructing the object's surface mesh; the volume-to-projected area ratio is obtained by dividing the object's volume by its projected area; and the continuity of grayscale changes along the Z-direction is obtained by counting the continuity of the slope of the grayscale curves of the object at each layer in the Z-direction. These three-dimensional combined features serve as input features for subsequent real particle discrimination and category recognition.

[0043] S303. The image analysis unit performs authenticity determination on each independent protrusion object based on the three-dimensional combined features extracted in step S302. Preferably, the authenticity determination adopts a two-level approach combining a rule engine and a deep learning discrimination model (such as an architecture based on U-Net, Mask R-CNN, etc.).

[0044] First, the rule engine performs an initial screening of independent protrusions according to preset discrimination rules: when the maximum height is less than the minimum effective height threshold, it is judged as an invalid micro-protrusion; when the contact boundary closure is lower than the first closure threshold and the object boundary and filter membrane texture transition continuously, it is judged as filter membrane burrs or background texture; when the surface roughness Sa and Sq are both lower than the smoothing threshold and the top curvature changes continuously, it is judged as liquid film residue or bubble-like interference objects based on the continuity of grayscale changes; when the aspect ratio exceeds the fiber abnormality threshold but the contact boundary is non-closed and stringy, it is judged as filter membrane fiber burrs; when the volume to projected area ratio is lower than the flat feature threshold and the grayscale change shows a specular reflection abnormal distribution, it is judged as an imaging artifact.

[0045] For independent protrusions that pass the initial screening by the rule engine, their corresponding local 3D voxel blocks, surface mesh data, and 3D combined feature vectors are then input into a trained deep learning discrimination model. The deep learning discrimination model outputs the probability values ​​of the object belonging to "real particles", "bubble residue", "dust adhesion", "filter membrane burrs", and "imaging artifacts".

[0046] Specifically, when the probability value corresponding to "real particle" is higher than the first confidence threshold, the object is determined to be a real particle; when the probability value is lower than the second confidence threshold, the object is determined to be an interference object; when the probability value is between the first confidence threshold and the second confidence threshold, the object is marked as a low-confidence object, and the process proceeds to step S305 to perform local verification.

[0047] S304. For objects identified as real particles in step S303, the image analysis unit further performs category recognition. Specifically, the three-dimensional combined feature vector of the real particle is input into the trained classification model to obtain the category to which the real particle belongs.

[0048] In this embodiment, the classification model (such as a Convolutional Neural Network (CNN) or a Support Vector Machine (SVM) outputs at least category labels such as crystal particles, fiber particles, metal particles, polymer particles, and other foreign matter particles. Specifically, when an object's surface has multiple facet transitions, a multi-peaked top curvature distribution, and uneven surface roughness, it is tended to be classified as a crystal particle; when the ratio of the object's length to width and length to height are both greater than a preset fiber threshold and extend continuously along the principal axis, it is tended to be classified as a fiber particle; when the object's surface is dense, has clear boundaries, high local reflection intensity, and low roughness, it is tended to be classified as a metal particle or a hard foreign matter particle. Finally, a category label and corresponding category confidence score are written for each real particle.

[0049] S305. For independent protruding objects marked as low-confidence objects in step S303, the central control unit establishes a local rescan task based on their center coordinates and external dimensions in the device coordinate system.

[0050] During local rescanning, the central control unit controls the high-precision XYZ 3D scanning stage to return to the area where the low-confidence object is located, and establishes a local scanning window centered on the center coordinates of the low-confidence object; then it switches to a high-magnification objective lens, and reduces the X-direction sampling interval, Y-direction sampling interval and Z-direction step size to below the original fine scanning parameters, and performs local high-density 3D slice acquisition around the low-confidence object, preferably acquiring slice data from the top, side walls and the boundary area in contact with the filter membrane of the object.

[0051] The image analysis unit performs local 3D reconstruction on the local rescanned data and recalculates the 3D combined features of the low-confidence object. It then calls the rule engine and deep learning discrimination model again to perform a re-determination of authenticity. If the re-determination result meets the condition of a real particle, it is included in the set of real particles; otherwise, it is recorded as an interference object.

[0052] Through the above-mentioned local rescanning and re-judgment process, low-confidence objects that were identified in the initial scan due to boundary occlusion, incomplete shape, or insufficient focal plane information can be supplemented for identification.

[0053] S4. Statistical Output of Detection Results S401. For the final determined set of real particles, the image analysis unit performs statistical caliber mapping according to preset detection standards. Specifically, for blocky real particles, the maximum projected size or equivalent volume diameter is preferably used as the standard particle size characterization value; for fibrous real particles, the length parameter is preferably used as the main statistical caliber; for multiple objects that are in contact with each other but have multiple independent height peaks in three-dimensional space and can be separated by connected component splitting, the real particles after separation are counted separately; for real particles partially embedded in the filter membrane pores, the exposed size and overall size can be calculated based on the morphology of the exposed part and the overall reconstructed morphology, and the standard statistical size is selected according to preset rules. Subsequently, each real particle is divided into the corresponding particle size interval according to the standard particle size characterization value, and the number of particles in each particle size interval, the number of particles of each category, and the unit volume concentration value are counted respectively.

[0054] S402. The central control unit automatically generates a test report based on the statistical results obtained in step S401. The test report includes at least the sample number, test time, test volume, filter membrane number, particle size distribution results, number of particles of each category, representative real particle three-dimensional morphology images, typical particle parameter table, and test conclusions. Simultaneously, the original optical slice data, the continuous three-dimensional dataset of the entire membrane, the independent protruding object mask, the authenticity judgment results, the category identification results, and the statistical results are written into the database for subsequent traceability, verification, and model updates.

[0055] The above process, through the construction of a full-process detection mechanism including sample loading and filter membrane carrier preparation, confocal 3D data acquisition of the filter membrane surface, intelligent particle analysis based on 3D morphology, and statistical output of detection results, can form a particle carrier suitable for stable imaging in the liquid sample to be tested. Based on the local 3D reference surface of the filter membrane, it distinguishes between the filter membrane background structure and the particle object. On this basis, it extracts the 3D combination features of independent protrusion objects, thereby achieving accurate discrimination between real particles and interference objects, identification of the category of real particles, and statistical output under standard caliber. Therefore, it can not only improve the accuracy of insoluble particle extraction and identification under complex filter membrane background, but also improve the efficiency of 3D data acquisition, the completeness of detection results, and the reliability of statistical output, thereby realizing the automation, intelligence, and traceability of the insoluble particle detection process.

[0056] Example 2 like Figure 2 As shown, an intelligent particle detection system based on laser confocal three-dimensional imaging is used for the automatic detection of insoluble particles in liquid samples. The system includes a central control unit, a sample preparation unit, a confocal imaging unit, and an image analysis unit. The units are interconnected and work together under the unified scheduling of the central control unit to complete sample processing, three-dimensional imaging, particle identification and analysis, and output of detection results.

[0057] The central control unit is used to establish detection tasks, receive sample information and detection parameters, generate control commands, and coordinate the sample preparation unit, confocal imaging unit, and image analysis unit to work together. Preferably, the central control unit also has a human-computer interaction function, used to input sample number, detection volume, filter membrane specifications, detection standards, and scanning parameters, and to display and output the detection results.

[0058] The sample preparation unit is used to acquire the liquid sample to be tested, filter the liquid sample to be tested, so that insoluble particles are retained on the surface of the filter membrane and / or the pore area of ​​the filter membrane, and dry, transport and position the filter membrane. Preferably, the sample preparation unit includes an automatic sample introduction component, a fluid control component, a filtration component, a drying component and a filter membrane transport component, wherein the fluid control component is used to transport the liquid sample to be tested to the filtration component, the filtration component is used to complete the particle retention, the drying component is used to dry the filter membrane, and the filter membrane transport component is used to transport the filter membrane to the loading area of ​​the confocal imaging unit.

[0059] The confocal imaging unit is used to pre-scan the reference area on the filter membrane and perform coarse scanning, adaptive fine scanning, and local rescanning on the entire filter membrane to obtain continuous three-dimensional data of the filter membrane surface. The confocal imaging unit includes a laser source, a confocal optical path system, an autofocus system, and a high-precision XYZ three-dimensional scanning stage. The laser source provides the scanning beam, the confocal optical path system acquires the optical signal corresponding to the focal plane, the autofocus system determines the starting focal plane for scanning, and the high-precision XYZ three-dimensional scanning stage carries and moves the filter membrane.

[0060] The image analysis unit is used to extract independent protrusion objects from continuous three-dimensional data based on the local three-dimensional reference surface of the filter membrane, extract three-dimensional combined features from the independent protrusion objects, and perform discrimination between real particles and interference objects, identification of real particle categories, and statistical caliber mapping.

[0061] Specifically, the image analysis unit includes a reference surface construction module, a 3D reconstruction module, an object extraction module, a feature extraction module, a discrimination module, a category recognition module, and a statistical mapping module. The reference surface construction module is used to establish a local 3D reference surface for the filter membrane, the 3D reconstruction module is used to form continuous 3D data, the object extraction module is used to extract independent protruding objects, the feature extraction module is used to extract the 3D combined features of objects, the discrimination module is used to perform discrimination between real particles and interference objects, the category recognition module is used to perform real particle category recognition, and the statistical mapping module is used to output statistical results according to preset detection standards.

[0062] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A smart particle detection method based on laser confocal three-dimensional imaging, characterized in that, Includes the following steps: S1. Sample loading and filter membrane carrier preparation: Obtain the liquid sample to be tested, filter the liquid sample to be tested so that insoluble particles are trapped on the surface of the filter membrane and / or the pore area of ​​the filter membrane, dry, transport and position the filter membrane to establish the initial spatial coordinates of the filter membrane relative to the equipment coordinate system; S2. Confocal 3D Data Acquisition of Filter Membrane Surface: A pre-scan is performed on the reference area on the filter membrane, and a local 3D reference surface of the filter membrane is established based on the 3D data of the reference area; on the basis of the local 3D reference surface of the filter membrane, a coarse scan is performed on the entire filter membrane to obtain the regional distribution results, and then an adaptive confocal fine scan is performed on different regions according to the regional distribution results to obtain a continuous 3D dataset of the filter membrane surface. S3. Intelligent particle analysis based on three-dimensional morphology: The continuous three-dimensional dataset is mapped to the corresponding local three-dimensional reference surface of the filter membrane. Candidate foreground voxels are extracted based on the difference between the voxel height value and the reference surface height value. Independent protrusion objects are extracted through three-dimensional connected domain analysis, outer envelope morphology analysis, boundary closure analysis and contact relationship analysis. Three-dimensional combined features are extracted from the independent protrusion objects, and the independent protrusion objects are distinguished as real particles and interference objects, and the real particle categories are identified based on the three-dimensional combined features; wherein, for independent protrusion objects with low confidence in the discrimination result, a local rescan is performed, and the discrimination is re-performed based on the rescan data; S4. Statistical Output of Detection Results: Based on the preset detection standards, the three-dimensional parameters of real particles are mapped to statistical caliber parameters, the real particles are classified and statistically analyzed, and the detection results are output.

2. The intelligent particle detection method based on laser confocal three-dimensional imaging according to claim 1, characterized in that, In step S1, a constant flow filtration method is used when filtering the liquid sample to be tested, and after filtration, a clean and dry airflow is introduced into the surface of the filter membrane to gently dry the filter membrane until the continuous liquid film on the surface of the filter membrane disappears or the thickness of the residual liquid film is lower than a preset threshold.

3. The intelligent particle detection method based on laser confocal three-dimensional imaging according to claim 1, characterized in that, The establishment of the local three-dimensional reference surface of the filter membrane includes: The reference area's 3D data is subjected to noise reduction, background correction, and intensity equalization. The reference area is divided into multiple local windows according to the preset window size; In each local window, a set of surface points on the exposed filter membrane surface is extracted, and a local reference surface for the corresponding local window is obtained by fitting the set of surface points. Based on the reference height value, local roughness reference value, and curvature variation range of each local reference surface, adjacent local reference surfaces are spliced ​​together to form a local three-dimensional reference surface for the filter membrane.

4. The intelligent particle detection method based on laser confocal three-dimensional imaging according to claim 1, characterized in that, The adaptive confocal fine scanning includes: Based on the height variation range, brightness variation range and suspected protrusion distribution density obtained from the coarse scanning of the filter membrane, the scanning area is divided into flat area, gradually changing area, high undulation area and suspected object dense area. For flat areas, fine scanning is performed using the first magnification and the first Z-axis step size; for gradually changing areas, fine scanning is performed using the first magnification and the second Z-axis step size; for areas with high undulations, fine scanning is performed using the second magnification and the third Z-axis step size; and for areas with dense suspected objects, fine scanning is performed using the second magnification and the scanning range is reduced. Among them, the second Z-axis step size is smaller than the first Z-axis step size, and the third Z-axis step size is smaller than the second Z-axis step size.

5. The intelligent particle detection method based on laser confocal three-dimensional imaging according to claim 1, characterized in that, The independent protrusion object satisfies the following conditions: at least a portion of the voxel height difference is greater than the object height threshold, an independent protrusion envelope is formed, and there is an identifiable contact boundary between it and the local three-dimensional reference surface of the filter membrane; In addition, for all voxel height differences within the local roughness reference value range, or for local connected regions where there are local protrusions but the boundaries are continuously transitioned with the background and do not form an independent envelope, they are judged as filter membrane surface texture or filter membrane fiber undulation; for regions that are lower than the local reference surface as a whole, they are judged as filter membrane pores, depressions or membrane surface defects.

6. The intelligent particle detection method based on laser confocal three-dimensional imaging according to claim 1, characterized in that, The three-dimensional composite features include at least the maximum projected size, maximum height, object volume, true surface area, aspect ratio, sphericity, top curvature distribution, surface roughness, contact boundary closure with the local three-dimensional reference surface of the filter membrane, volume to projected area ratio, and grayscale variation continuity along the Z direction.

7. The intelligent particle detection method based on laser confocal three-dimensional imaging according to claim 1, characterized in that, The distinction between real particles and interference objects adopts a two-level discrimination method that combines a rule engine and a deep learning discrimination model. Among them, the rule engine performs initial screening of independent protrusion objects based on maximum height, contact boundary closure, surface roughness, aspect ratio, volume to projected area ratio, and the continuity of grayscale change along the Z direction. The deep learning discrimination model outputs discrimination results for real particles, bubble residues, dust adhesion, filter membrane burrs, or imaging artifacts based on the local three-dimensional voxel blocks, surface mesh data, and three-dimensional combined feature vectors corresponding to the independent protruding objects after initial screening.

8. The intelligent particle detection method based on laser confocal three-dimensional imaging according to claim 1, characterized in that, The local rescanning includes: establishing a local scanning window based on the center coordinates and circumscribed dimensions of the low-confidence independent protrusion in the device coordinate system; switching to a high-magnification objective lens; reducing the X-axis sampling interval, Y-axis sampling interval, and Z-axis step size to below the original fine scanning parameters; acquiring additional three-dimensional slice data of the top, sidewalls, and contact boundary area with the filter membrane around the low-confidence independent protrusion; and then re-judging based on the acquired data.

9. An intelligent particle detection system based on laser confocal three-dimensional imaging, characterized in that, include: The central control unit is used to establish detection tasks, generate control commands, and schedule the various functional units to work together. The sample preparation unit is used to acquire the liquid sample to be tested, filter the liquid sample to be tested so that insoluble particles are trapped on the surface of the filter membrane and / or the pore area of ​​the filter membrane, and dry, transport and position the filter membrane. The confocal imaging unit is used to pre-scan the reference area on the filter membrane. Based on the pre-scan results of the reference area, it helps to establish a local three-dimensional reference surface of the filter membrane. It also performs coarse scanning, adaptive fine scanning and local rescanning on the entire filter membrane to obtain continuous three-dimensional data of the filter membrane surface. The image analysis unit is used to extract independent protrusion objects from continuous three-dimensional data based on the local three-dimensional reference surface of the filter membrane, extract three-dimensional combined features from the independent protrusion objects, and perform discrimination between real particles and interference objects, identification of real particle categories, and statistical caliber mapping. The central control unit, sample preparation unit, confocal imaging unit, and image analysis unit are interconnected to execute the intelligent particle detection method based on laser confocal three-dimensional imaging as described in any one of claims 1 to 8.

10. The intelligent particle detection system based on laser confocal three-dimensional imaging according to claim 9, characterized in that, The confocal imaging unit includes a laser source, a confocal optical path system, an autofocus system, and a high-precision XYZ three-dimensional scanning stage; The image analysis unit is configured to: map continuous three-dimensional data to the corresponding local three-dimensional reference surface of the filter membrane, extract independent protrusion objects based on height difference calculation and three-dimensional connected component analysis; and perform real particle and interference object discrimination on the independent protrusion objects based on the maximum projected size, maximum height, object volume, real surface area, aspect ratio, sphericity, top curvature distribution, surface roughness, contact boundary closure, volume to projected area ratio, and grayscale change continuity along the Z direction.

Images