A water quality micro-plastic rapid detection method based on multi-scale feature fusion

Abstract

The application discloses a kind of water quality microplastic rapid detection methods based on multi-scale feature fusion, comprising the following steps: obtaining water sample image and corresponding sampling volume, into spectrum-morphology coupling RepPoints network;RepPoints point set expression layer is reformed, and point set update is executed according to inner domain point, contour point, skeleton point;With the convergence range of inner domain point, the contour point offset boundary is limited;With skeleton point, the arrangement of contour point is corrected;Through spectrum-morphology consistency offset gate, regression direction is reserved or reverse attenuation is executed;Microplastic discrimination, particle counting, particle size determination and fiber morphology distinction are executed;Water microplastic concentration detection result is obtained by particle size-concentration inversion output head.The application realizes water quality microplastic rapid identification and concentration quantification, and improves detection precision and anti-interference ability.

Description

Technical Field

[0001] This invention relates to the field of water quality testing technology, and in particular to a rapid detection method for microplastics in water based on multi-scale feature fusion. Background Technology

[0002] Microplastic detection in water is used for environmental monitoring, water quality safety assessment, and pollution source analysis. Current detection methods mostly employ manual microscopic observation, fluorescence microscopy, multispectral imaging, image segmentation, and particle classification algorithms to identify suspected microplastic particles in water samples. Some methods then calculate the concentration based on particle number, particle size distribution, and sampling volume.

[0003] Existing image detection methods mostly rely on single features such as color, brightness, edge, and texture. When faced with interference from mud, organic debris, bubbles, and non-plastic fibers, they easily identify impurities as microplastic particles. Microplastic particles vary greatly in morphology; the boundaries of flaky, fragmented, and fibrous particles are irregular. Conventional detection frames cannot accurately fit the outer edge of the particles, and the central trend and lateral boundaries of fibrous microplastics are also difficult to represent stably, limiting the accuracy of particle size calculation.

[0004] Existing representative point set detection methods can replace rectangular detection boxes with several representative points to represent the spatial range of irregular targets. However, these representative points often participate in regression as similar positioning points and do not differentiate between interior points, contour points, and skeleton points for different particle morphologies in water microplastics. This makes it difficult to form a coordinated constraint between the particle body, particle outer edge, and the center line of slender particles. Existing multi-scale identification methods mostly remain at the level of feature splicing or classification output, lacking bidirectional constraints between spectral response and particle morphology. The connection between identification results, particle scale, fiber morphology, and sampling volume is insufficient, affecting the anti-interference ability, particle size statistical stability, and concentration quantification efficiency of rapid microplastic detection.

[0005] Therefore, how to provide a rapid detection method for water microplastics based on multi-scale feature fusion is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a rapid detection method for microplastics in water quality based on multi-scale feature fusion. This invention utilizes a cascaded heterogeneous point set layer to represent microplastic particles as interior points, contour points, and skeleton points. Point set offset is corrected by a spectral-morphology consistency offset gate, and concentration quantification is achieved through a particle size-concentration inversion output head. This method has the advantages of strong anti-interference, accurate particle size calculation, and high detection efficiency.

[0007] A rapid detection method for water microplastics based on multi-scale feature fusion according to an embodiment of the present invention includes the following steps: S1. Obtain the water sample image and corresponding sampling volume, and send the water sample image into the spectral-morphological coupling RepPoints network to perform cross-coding of the spectral response channel and the particle morphology texture channel. S2. Transform the RepPoints point set expression layer into a cascaded heterogeneous point set layer, and perform point set updates in the order of inner domain points, contour points, and skeleton points, so that the inner domain points first converge to the stable region of the particle body. S3. Limit the offset boundary of the contour point by the convergence range of the inner domain point, so that the contour point performs edge regression along the outer edge of the particle body, and restricts the contour point from expanding outward to the edge of the impurity. S4. Embed the skeleton points into the contour point regression process, so that the skeleton points extend along the center line of the slender particles and then the contour point arrangement is corrected in the reverse direction to complete the morphological constraint of the fibrous microplastics. S5. After the cascaded heterogeneous point set layer, a spectral-morphological consistency offset gate is set to retain the regression direction for the three types of points with consistent spectral response and continuous morphology, and to perform reverse attenuation for the three types of points with discrete spectral response and broken morphology. S6. Microplastic discrimination, particle counting, particle size determination and fiber morphology differentiation are performed using the inner domain points, contour points and skeleton points after spectral-morphological consistency offset gate correction. S7. Connect the inner domain points, contour points, and skeleton points after spectral-morphology consistency offset gate correction to the particle size-concentration inversion output head, and convert the water microplastic concentration detection results according to particle count, particle size, fiber morphology, and sampling volume.

[0008] Optionally, the spectral-morphology coupled RepPoints network in S1 includes a spectral response channel, a particle morphology texture channel, a cross-coding layer, a cascaded heterogeneous point set layer obtained by modifying the RepPoints point set expression layer, a spectral-morphology consistency shift gate, and a particle size-concentration inversion output head. The spectral response channel and the particle morphology texture channel are connected to the cross-coding layer, the cross-coding layer is connected to the cascaded heterogeneous point set layer, the cascaded heterogeneous point set layer is connected to the spectral-morphology consistency shift gate, and the spectral-morphology consistency shift gate is connected to the particle size-concentration inversion output head.

[0009] Optionally, the cross-coding in S1 specifically includes: By compressing adjacent spectral responses within a spectral response channel through one-dimensional channel convolution, spectral gate weights are formed. These spectral gate weights are then multiplied into the edge convolution path of the grain morphology texture channel to obtain the edge convolution encoding updated along the stable position of the spectral response. The boundary orientation and skeleton orientation within the particle morphology texture channel are extracted by directional convolution to form morphology gating weights. The morphology gating weights are multiplied into the channel convolution path of the spectral response channel to obtain the channel convolution encoding updated along the continuous position of the morphology. The edge convolutional encoding and channel convolutional encoding are added position by position; The code obtained by adding position by position is written back to the spectral response channel and the grain morphology texture channel respectively through residual connection, thus completing the cross-coding of the spectral response channel and the grain morphology texture channel.

[0010] Optionally, the point set update process in S2 specifically includes: Using the inner domain points as the first update points, the inner domain points are moved according to the direction from the low response position to the high response position in the encoding output of the cross-coding layer. The inner domain points that deviate from the stable region of the particle are removed, and the remaining inner domain points form the convergence range of the inner domain points. Use the contour points as the second-level update points to bind the convergence range of the contour points to the interior points. Use the skeleton point as the third-level update point to bind the skeleton point to the contour point corresponding to the same particle; Maintain the particle number binding relationship between the inland points, contour points, and skeleton points of the same particle, and output the inland points, contour points, and skeleton points as point sets of the cascaded heterogeneous point set layer.

[0011] Optionally, in S3, the contour point offset boundary is defined by the convergence range of the interior points, specifically including: Using the convergence range of the inner domain points as the offset boundary of the contour points, the edge response direction of the encoding output of the cross-coding layer near the outer edge of the convergence range of the inner domain points is read. Perform edge-fitting regression on the contour points along the edge response direction toward the outer edge of the particle body; When the contour point crosses the outer edge of the offset boundary, project the contour point to the nearest position on the offset boundary and press the contour point back along the direction pointing to the outer edge of the particle body. When the contour point is located inside the offset boundary and fits the outer edge of the particle body, the contour point position is retained.

[0012] Optionally, the process of embedding skeleton points into contour point regression in S4 specifically includes: The skeleton point reads the directional continuous response in the output of the cross-coding layer and moves along the center line of the slender particle. The direction of the line connecting adjacent skeleton points is used as the direction of skeleton point movement and written into the contour point edge-fitting regression process. Using the direction of skeleton point movement as an axial reference, the contour points on both sides of the slender particles are grouped and paired. Based on the distance difference between the paired contour points and the corresponding skeleton points, adjust the arrangement order and distance between the contour points on both sides to keep the contour points in an arrangement that corresponds to the shape of the slender particles.

[0013] Optionally, the spectral-morphological consistency shift gate processing in S5 specifically includes: The inner domain points, contour points, and skeleton points in the output of the cascaded heterogeneous point set layer are input into the spectral-morphological consistency offset gate, so that the inner domain points enter the main response gated branch, the contour points enter the boundary continuity gated branch, and the skeleton points enter the directional extension gated branch. In the main response gating branch, the spectral response code is read along the regression direction of the inner domain point. When the spectral response is consistent, the regression direction of the inner domain point is retained. When the spectral response is discrete, the offset amplitude of the inner domain point is shortened. In the boundary continuous gating branch, the particle morphology texture code is read along the regression direction of the contour point. When the morphology is continuous, the regression direction of the contour point is retained. When the morphology is broken, the contour point is pressed back to the outer edge of the particle body. In the directional extension gated branch, the particle morphology texture code is read along the skeleton point movement direction. When the morphology is continuous, the skeleton point movement direction is retained. When the morphology is broken, the skeleton point movement direction is reversed and attenuated. The inland points processed by the main response gating branch, the contour points processed by the boundary continuous gating branch, and the skeleton points processed by the direction extension gating branch are output according to the particle number binding relationship to obtain the inland points, contour points, and skeleton points after spectral-morphological consistency offset gate correction.

[0014] Optionally, the process of microplastic identification, particle counting, particle size determination, and fiber morphology differentiation in S6 specifically includes: The particle body is determined by the inner domain points after spectral-morphological consistency offset gate correction. When the inner domain points are located in the stable region of the particle body and the contour points and skeleton points correspond to the inner domain points according to the particle number binding relationship, the corresponding particles are identified as microplastic particles. Determine the inner domain point, outline point, and skeleton point corresponding to each microplastic particle according to the particle number binding relationship, and accumulate the number of microplastic particles according to each particle number to complete the particle counting. Based on the particle number binding relationship, select the contour points corresponding to the same particle, take the distance between the two contour points with the largest distance as the major axis scale, and take the maximum span of the contour points perpendicular to the major axis scale as the minor axis scale. According to the particle number binding relationship, select the skeleton point, major axis scale and minor axis scale corresponding to the same particle. When the skeleton point is continuously distributed along the center line of the slender particle and the ratio of the major axis scale to the minor axis scale reaches the set aspect ratio threshold, the corresponding microplastic particles are classified into fiber morphology. According to the particle number binding relationship, select the skeleton point corresponding to the same particle. When the skeleton point is not continuously distributed along the center line of the slender particle, the corresponding microplastic particle is classified as non-fiber particle. Based on the particle number binding relationship, the major axis and minor axis dimensions corresponding to the same particle are selected. When the ratio of the major axis dimension to the minor axis dimension does not reach the set aspect ratio threshold, the corresponding microplastic particle is classified as a non-fiber particle.

[0015] Optionally, the processing of the particle size-concentration inversion output head in S7 specifically includes: Receive the interior points, contour points, and skeleton points after spectral-morphological consistency offset gate correction, call the sampling volume, and call the major axis scale, minor axis scale, fiber morphology and non-fiber particles corresponding to each microplastic particle according to the particle number binding relationship. For microplastic particles classified into fibrous morphology, the distance between adjacent skeleton points is accumulated along the direction of skeleton point movement, and the cumulative result is used to correct the major axis scale. The corrected major axis scale is then used as the particle scale. For microplastic particles that are classified as non-fibrous particles, the particle size is determined using the long axis and short axis dimensions. For microplastic particles that retain the regression direction through the main response gated branch and whose skeleton points under the same particle number do not undergo reverse attenuation through the directional extension gated branch, the full count weight is assigned; for microplastic particles that undergo reverse attenuation through the directional extension gated branch, the reduced count weight is assigned. The counting weight of each microplastic particle is assigned to the corresponding particle size interval according to the particle size. The counting weights in each particle size interval are summed to obtain the weighted number of particles in each particle size interval. The weighted number of particles in each particle size range is divided by the sampling volume to obtain the microplastic concentration in each particle size range. The microplastic concentrations in each particle size range are summed to obtain the microplastic concentration detection results in the water body.

[0016] The beneficial effects of this invention are: (1) This invention transforms the RepPoints point set expression layer into a cascaded heterogeneous point set layer, so that the inner domain points, contour points and skeleton points participate in the particle body positioning, particle outer edge fitting and slender particle center line characterization, respectively, avoiding the rough enclosure of irregular microplastic particle boundaries by conventional detection boxes, and improving the boundary expression accuracy of fragmented, sheet-like and fibrous microplastics.

[0017] (2) The present invention limits the boundary of the contour point offset by the convergence range of the inner domain point, so that the contour point performs edge regression along the outer edge of the particle body, and restricts the contour point from expanding outward to the edge of silt, organic debris, bubbles and non-plastic fibers; the arrangement of the contour points on both sides of the slender particle is corrected by the movement direction of the skeleton point, thereby improving the stability of microplastic identification and the accuracy of fiber morphology differentiation in complex water samples.

[0018] (3) The present invention corrects the regression direction and offset amplitude of the inland points, contour points and skeleton points by using the spectral-morphology consistency offset gate, so that the points with discrete spectral response and broken morphology are subjected to reverse attenuation, reducing the impact of abnormal points on particle size and concentration conversion; the corrected inland points, contour points and skeleton points are connected to the particle size-concentration inversion output head, so that microplastic discrimination, particle counting, particle size determination, fiber morphology differentiation and concentration conversion form a continuous processing chain, improving the detection efficiency of microplastics in water and the quantitative accuracy of concentration. Attached Figure Description

[0019] 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 a rapid detection method for water microplastics based on multi-scale feature fusion proposed in this invention; Figure 2 This is a schematic diagram of the spectral-morphological coupled RepPoints network structure proposed in this invention; Figure 3 This is a schematic diagram of the cascaded heterogeneous point set layer and spectral-morphological consistency shift gate processing proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figures 1-3 A rapid detection method for water microplastics based on multi-scale feature fusion includes the following steps: S1. Obtain the water sample image and corresponding sampling volume, and send the water sample image into the spectral-morphological coupling RepPoints network to perform cross-coding of the spectral response channel and the particle morphology texture channel. S2. Transform the RepPoints point set expression layer into a cascaded heterogeneous point set layer, and perform point set updates in the order of inner domain points, contour points, and skeleton points, so that the inner domain points first converge to the stable region of the particle body. S3. Limit the offset boundary of the contour point by the convergence range of the inner domain point, so that the contour point performs edge regression along the outer edge of the particle body, and restricts the contour point from expanding outward to the edge of the impurity. S4. Embed the skeleton points into the contour point regression process, so that the skeleton points extend along the center line of the slender particles and then the contour point arrangement is corrected in the reverse direction to complete the morphological constraint of the fibrous microplastics. S5. After the cascaded heterogeneous point set layer, a spectral-morphological consistency offset gate is set to retain the regression direction for the three types of points with consistent spectral response and continuous morphology, and to perform reverse attenuation for the three types of points with discrete spectral response and broken morphology. S6. Microplastic discrimination, particle counting, particle size determination and fiber morphology differentiation are performed using the inner domain points, contour points and skeleton points after spectral-morphological consistency offset gate correction. S7. Connect the inner domain points, contour points, and skeleton points after spectral-morphology consistency offset gate correction to the particle size-concentration inversion output head, and convert the water microplastic concentration detection results according to particle count, particle size, fiber morphology, and sampling volume.

[0022] In this invention, the particle size-concentration inversion output head receives the inner domain points, contour points, and skeleton points after spectral-morphology consistency offset gate correction; it reads the major axis scale, minor axis scale, fiber morphology, and non-fiber particles corresponding to each microplastic particle one by one according to the particle number binding relationship; the particle size-concentration inversion output head completes particle size conversion, count weight assignment, and particle size interval classification under the same particle number; the sampling volume is used as a concentration normalization parameter to participate in the final conversion.

[0023] In this embodiment, the spectral-morphology coupled RepPoints network in S1 includes a spectral response channel, a particle morphology texture channel, a cross-coding layer, a cascaded heterogeneous point set layer obtained by modifying the RepPoints point set expression layer, a spectral-morphology consistency shift gate, and a particle size-concentration inversion output head. The spectral response channel and the particle morphology texture channel are connected to the cross-coding layer, the cross-coding layer is connected to the cascaded heterogeneous point set layer, the cascaded heterogeneous point set layer is connected to the spectral-morphology consistency shift gate, and the spectral-morphology consistency shift gate is connected to the particle size-concentration inversion output head.

[0024] In a specific embodiment of the present invention, the spectral-morphological coupled RepPoints network structure specifically includes: The spectral response channel and the particle morphology texture channel are set side-by-side at the network entry point. The spectral response channel receives spectral response information from the water sample image and encodes variations in fluorescence response intensity, differences in multispectral channels, and response differences between particle regions and background regions. The particle morphology texture channel receives particle morphology texture information from the water sample image and encodes particle edges, particle texture orientation, and particle skeleton trends.

[0025] Both the spectral response channel and the grain morphology texture channel are connected to the cross-coding layer. The cross-coding layer receives the spectral response code output from the spectral response channel and the grain morphology texture code output from the grain morphology texture channel. It writes the spectral response code into the grain morphology texture coding process and writes the grain morphology texture code into the spectral response code process, thus enabling the two channels to be bidirectionally constrained at the same spatial location.

[0026] The cross-coding layer connects to the cascaded heterogeneous point set layer. The cascaded heterogeneous point set layer receives the cross-coding results from the cross-coding layer and, based on the representative points of the RepPoints point set expression layer, sets interior points, contour points, and skeleton points, transforming the general representative point expression into a cascaded heterogeneous point set expression of interior points, contour points, and skeleton points. The cascaded heterogeneous point set layer updates the point set in the order of interior points, contour points, and skeleton points, enabling interior points to participate in particle body localization, contour points to participate in particle boundary fitting, and skeleton points to participate in the characterization of the centerline of slender particles.

[0027] A cascaded heterogeneous point set layer is connected to a spectral-morphology consistency offset gate. The spectral-morphology consistency offset gate receives interior points, contour points, and skeleton points output from the cascaded heterogeneous point set layer and corrects the offset direction, offset amplitude, and point participation weights of these points. Based on spectral response consistency, morphological continuity, and background contamination levels, the spectral-morphology consistency offset gate enhances the offset amplitude retention ratio for interior points, contour points, and skeleton points located in the stable response region of microplastics, and reduces the participation weight for interior points, contour points, and skeleton points located in impurity interference regions.

[0028] The spectral-morphology consistency offset gate connects to the particle size-concentration inversion output head. The particle size-concentration inversion output head receives the interior points, contour points, and skeleton points corrected by the spectral-morphology consistency offset gate. Based on the interior points, the output head performs particle counting; based on the contour points, it determines the particle size; and based on the skeleton points, it corrects the fiber morphology. Finally, it combines the sample volume to calculate the water microplastic concentration detection results.

[0029] In the data stream of the spectral-morphology coupled RepPoints network, the water sample image first enters the spectral response channel and the particle morphology texture channel; the encoding results of the two channels enter the cross-coding layer; the output of the cross-coding layer enters the cascaded heterogeneous point set layer; the interior points, contour points, and skeleton points output by the cascaded heterogeneous point set layer enter the spectral-morphology consistency offset gate; the interior points, contour points, and skeleton points corrected by the spectral-morphology consistency offset gate enter the particle size-concentration inversion output head; the particle size-concentration inversion output head outputs the water microplastic concentration detection results.

[0030] The aforementioned network structure forms a continuous processing chain consisting of the spectral response channel, particle morphology and texture channel, cross-coding layer, cascaded heterogeneous point set layer, spectral-morphology consistency offset gate, and particle size-concentration inversion output head. The spectral response channel and particle morphology and texture channel do not output detection results independently; the cross-coding layer does not output classification results independently; the cascaded heterogeneous point set layer does not output concentration results independently; the spectral-morphology consistency offset gate does not change the names of interior points, contour points, and skeleton points; and the particle size-concentration inversion output head uses the point set correction results and sampling volume from the same processing chain to complete the concentration conversion.

[0031] In this embodiment, the cross-coding in S1 specifically includes: By compressing adjacent spectral responses within a spectral response channel through one-dimensional channel convolution, spectral gate weights are formed. These spectral gate weights are then multiplied into the edge convolution path of the grain morphology texture channel to obtain the edge convolution encoding updated along the stable position of the spectral response. The process of forming spectral gate weights through one-dimensional channel convolution in this invention specifically includes: arranging adjacent spectral responses along the channel dimension of the spectral response channel; performing sliding compression on adjacent spectral responses using one-dimensional channel convolution; and normalizing and mapping the compression result to obtain the spectral gate weights corresponding to the spatial position of the particle morphology texture channel. The boundary orientation and skeleton orientation within the particle morphology texture channel are extracted by directional convolution to form morphology gating weights. The morphology gating weights are multiplied into the channel convolution path of the spectral response channel to obtain the channel convolution encoding updated along the continuous position of the morphology. The process of forming morphology gating weights through directional convolution in this invention specifically includes: setting horizontal, vertical, and diagonal directional convolution kernels in the particle morphology texture channel; scanning the particle boundary region and skeleton region with the directional convolution kernels respectively; and normalizing the directional convolution response to obtain the morphology gating weights. In this invention, spectral gate weights are multiplied with features in the edge convolution path at the same spatial position; morphological gate weights are multiplied with features in the channel convolution path at the same spatial position; when the number of channels is different, channel mapping is used to make the number of channels consistent before multiplication; the results after multiplication are used as edge convolution codes and channel convolution codes, respectively. The edge convolutional encoding and channel convolutional encoding are added position by position; The position-by-position addition process in this invention specifically includes: first adjusting edge convolutional coding and channel convolutional coding to the same spatial size; then adjusting edge convolutional coding and channel convolutional coding to the same number of channels; performing addition at the same pixel position and the same channel position; and using the addition result as the fusion code after cross-coding. The code obtained by adding position by position is written back to the spectral response channel and the grain morphology texture channel respectively through residual connection, thus completing the cross-coding of the spectral response channel and the grain morphology texture channel.

[0032] The write-back process in this invention specifically includes: feeding the code obtained by adding position by position into two residual branches respectively; adding one residual branch to the original code of the spectral response channel to update the spectral response channel; adding the other residual branch to the original code of the particle morphology texture channel to update the particle morphology texture channel; and retaining the original channel names of the two channels after write-back without renaming them.

[0033] In this embodiment, the process of updating the point set in S2 specifically includes: Using the inner domain points as the first update points, the inner domain points are moved according to the direction from the low response position to the high response position in the encoding output of the cross-coding layer. The inner domain points that deviate from the stable region of the particle are removed, and the remaining inner domain points form the convergence range of the inner domain points. In this invention, the first update point refers to the point that performs the position update first in the cascaded heterogeneous point set layer. The inner domain point is the first update point, the contour point is the second update point, and the skeleton point is the third update point. The codes output by the cross-coding layer are arranged according to spatial location, with each spatial location corresponding to a coding vector; the cascaded heterogeneous point set layer reads the coding vector in the neighborhood of the inner domain point, uses the response intensity component in the coding vector for the movement of the inner domain point, uses the edge response component in the coding vector for the contour point regression, and uses the directional continuous component in the coding vector for the movement of the skeleton point. The direction confirmation specifically includes: setting a neighborhood window centered on each intra-domain point, comparing the response intensity of adjacent spatial positions within the neighborhood window; selecting the position with the lowest response intensity as the starting position and the position with the highest response intensity as the target position; using the vector from the starting position to the target position as the movement direction of the intra-domain point; and performing one or more position offsets according to this movement direction. In the encoding output of the cross-coding layer, the spatial location with continuous spectral response, continuous particle morphology and texture, and local response gradient convergence towards the center is selected as the stable region of the particle body; when the inner domain point moves and falls into this region, it is retained; when the inner domain point moves and enters the background noise region, the impurity edge region, or the response discrete region, it is discarded. The retained intra-domain points are sorted according to their spatial adjacency. Intra-domain points whose adjacent distance does not exceed the preset adjacency distance are connected, while those whose adjacent distance exceeds the preset adjacency distance are disconnected. The connected intra-domain points form a closed or semi-closed range, which serves as the convergence range of the intra-domain points. The preset adjacency distance is determined based on the average spacing of intra-domain points of the same particle, and is taken as 1.2 to 2.0 times the average spacing of intra-domain points. Use the contour points as the second-level update points to bind the convergence range of the contour points to the interior points. The binding operation in this invention specifically includes: associating the convergence range of contour points and interior points in the same particle with the same particle number; when performing edge regression on contour points, only reading the convergence range of interior points under the same particle number, and not reading the convergence range corresponding to other particles; Use the skeleton point as the third-level update point to bind the skeleton point to the contour point corresponding to the same particle; The specific operations to prevent the mixing of different particle point sets in this invention include: assigning the same particle number to the same particle in the cascaded heterogeneous point set layer, and writing the internal domain points, contour points, and skeleton points under the same particle number into the same point set group; the skeleton points only perform subsequent directional constraints on the contour points under the same particle number. Maintain the particle number binding relationship between the inland points, contour points, and skeleton points of the same particle, and output the inland points, contour points, and skeleton points as point sets of the cascaded heterogeneous point set layer.

[0034] In this embodiment, S3, which defines the contour point offset boundary by the convergence range of the inner domain points, specifically includes: Using the convergence range of the inner domain points as the offset boundary of the contour points, the edge response direction of the encoding output of the cross-coding layer near the outer edge of the convergence range of the inner domain points is read. In this invention, the connecting lines between adjacent inner domain points are extracted along the outer edge of the convergence range of the inner domain points, and the outer edge enclosed by the connecting lines is used as the offset boundary of the contour points. When the contour points are regressed, their movement position must not exceed the outer edge of the offset boundary. For each contour point, a neighborhood window is selected near the outer edge of the convergence range of the inner domain points. The encoding output by the cross-coding layer is read in the neighborhood window, and the edge response intensity of adjacent spatial positions is compared. The direction with the largest change in edge response intensity is taken as the edge response direction. Perform edge-fitting regression on the contour points along the edge response direction toward the outer edge of the particle body; The edge regression in this invention specifically includes: starting from the current spatial position, the contour point performs a position offset along the edge response direction; after offset, the edge response intensity is read again; when the edge response intensity at the offset position increases, the offset position is retained; when the edge response intensity at the offset position decreases, it reverts to the position before offset or the nearest position inside the offset boundary. When the contour point crosses the outer edge of the offset boundary, project the contour point to the nearest position on the offset boundary and press the contour point back along the direction pointing to the outer edge of the particle body. The pressing back operation in this invention specifically includes: determining the relationship between the position of the offset point and the inner and outer sides of the offset boundary; if the offset point is located outside the offset boundary, calculating the nearest projection position of the offset point to the offset boundary; moving the offset point to the nearest projection position, and then moving it in the direction pointing to the outer edge of the particle body, so that the offset point returns to the vicinity of the outer edge of the particle body. When the contour point is located inside the offset boundary and fits the outer edge of the particle body, the contour point position is retained.

[0035] The fitting determination conditions in this invention include: when the contour point is located inside the offset boundary and the edge response intensity in the neighborhood of the contour point reaches the edge response threshold, the contour point is determined to fit the outer edge of the particle body; the cascaded heterogeneous point set layer retains the position of the contour point.

[0036] In this embodiment, the process of embedding skeleton points into contour point regression in S4 specifically includes: The skeleton point reads the directional continuous response in the output of the cross-coding layer and moves along the center line of the slender particle. In this invention, the directional continuous component is read from the output of the cross-coding layer; the directional changes of adjacent spatial positions are compared in the neighborhood of the skeleton point; the direction with the smallest directional change and the largest extension length is taken as the direction corresponding to the directional continuous response; the cascaded heterogeneous point set layer searches for adjacent skeleton points along the extension direction of the directional continuous response in the elongated particle region; the direction of the line with the smallest directional change between adjacent skeleton points is taken as the direction of the center line of the elongated particle; the skeleton point moves its position along this direction. The direction of the line connecting adjacent skeleton points is used as the direction of skeleton point movement and written into the contour point edge-fitting regression process. In this invention, the skeleton points are arranged according to the projection order of the skeleton points on the center line direction of the slender particles; two skeleton points with adjacent projection order are connected; and the direction of the connecting line is used as the direction of movement of the skeleton points. During the contour point edge-fitting regression process, the contour points simultaneously read the edge response direction and the skeleton point movement direction. When the edge response direction and the skeleton point movement direction remain continuous, the contour points move along the edge response direction. When the edge response direction deviates from the skeleton point movement direction, the contour point movement direction is corrected according to the skeleton point movement direction. Using the direction of skeleton point movement as an axial reference, the contour points on both sides of the slender particles are grouped and paired. The grouping and pairing process in this invention includes: using the direction of skeleton point movement as an axial reference line, dividing the contour points located on both sides of the axial reference line into two groups; arranging the contour points in each group according to the projection order of the contour points in the direction of skeleton point movement; and pairing the contour points whose projection positions are adjacent and belong to the two sides. Based on the distance difference between the paired contour points and the corresponding skeleton points, adjust the arrangement order and distance between the contour points on both sides to keep the contour points in an arrangement that corresponds to the shape of the slender particles.

[0037] In this invention, for each pair of paired contour points, the skeleton point with the closest projection position is selected as the corresponding skeleton point; the distances from the two contour points to the corresponding skeleton point are calculated respectively, and the difference between the two is taken as the distance difference; When the distance difference exceeds the preset adjustment range, the contour point with the larger distance is moved in the inner direction perpendicular to the direction of skeleton point movement, and the contour point with the smaller distance is moved in the outer direction perpendicular to the direction of skeleton point movement; after moving, the distance difference is recalculated until the distance difference enters the preset adjustment range.

[0038] In this embodiment, the processing of the spectral-morphological consistency shift gate in S5 specifically includes: The inner domain points, contour points, and skeleton points in the output of the cascaded heterogeneous point set layer are input into the spectral-morphological consistency offset gate, so that the inner domain points enter the main response gated branch, the contour points enter the boundary continuity gated branch, and the skeleton points enter the directional extension gated branch. In a specific embodiment of the present invention, the main response gating branch receives the inner domain points in the point set output of the cascaded heterogeneous point set layer; for each inner domain point, a sampling window is established with the current coordinates of the inner domain point as the center, and forward and backward positions are selected within the sampling window along the regression direction of the inner domain point; the main response gating branch reads the spectral response codes corresponding to the forward position, the current coordinates of the inner domain point, and the backward position, and compares the spectral response changes at the three positions; when the spectral response changes at the three positions are lower than the set response difference threshold, it is determined that the spectral response is consistent, and the regression direction and offset amplitude of the inner domain point are retained; when the spectral response changes at the three positions reach the set response difference threshold, it is determined that the spectral response is discrete, and the offset amplitude of the inner domain point is shortened proportionally; the set response difference threshold is determined based on the spectral response fluctuation of the background area in the water sample image, and is taken as 1.5 times to 3 times the standard deviation of the background spectral response; The boundary continuity gated branch receives contour points from the output of the cascaded heterogeneous point set layer. For each contour point, it reads the particle morphology texture code of the adjacent position along the regression direction of the contour point and compares the boundary orientation change and texture orientation change of the adjacent position. The morphology difference threshold is set according to the particle boundary orientation change and skeleton orientation change, and is taken from 10 degrees to 25 degrees. When both the boundary orientation change and the texture orientation change are lower than the set morphology difference threshold, it is determined to be morphologically continuous and the regression direction of the contour point is retained. When at least one of the boundary orientation change and the texture orientation change reaches the set morphology difference threshold, it is determined to be morphologically broken. When the morphology is broken, the boundary continuity gated branch reads the offset boundary, projects the contour point to the nearest position on the offset boundary, and presses the contour point back along the direction pointing to the outer edge of the particle body. The directional extension gated branch receives skeleton points from the output of the cascaded heterogeneous point set layer. For each skeleton point, it reads the particle morphology texture code of the adjacent position along the skeleton point's movement direction and compares the skeleton orientation changes of the adjacent positions. When the skeleton orientation change is lower than the set morphological difference threshold, it is determined to be morphologically continuous, and the skeleton point's movement direction is retained. When the skeleton orientation change reaches the set morphological difference threshold, it is determined to be morphologically broken. When the morphologically broken, the directional extension gated branch reverses the skeleton point's movement direction and shortens the skeleton point's offset amplitude in the reversed direction by using a reverse attenuation coefficient. In the main response gating branch, the spectral response code is read along the regression direction of the inner domain point. When the spectral response is consistent, the regression direction of the inner domain point is retained. When the spectral response is discrete, the offset amplitude of the inner domain point is shortened. In this invention, when shortening the offset amplitude, the spectral-morphological consistency offset gate reads the offset of the inland point, contour point, and skeleton point before entering the gated branch, multiplies the offset by the offset shortening coefficient, and obtains the shortened offset; the offset shortening coefficient is greater than 0 and less than 1; the inland point, contour point, and skeleton point update their positions according to the shortened offset. In the boundary continuous gating branch, the particle morphology texture code is read along the regression direction of the contour point. When the morphology is continuous, the regression direction of the contour point is retained. When the morphology is broken, the contour point is pressed back to the outer edge of the particle body. In the directional extension gated branch, the particle morphology texture code is read along the skeleton point movement direction. When the morphology is continuous, the skeleton point movement direction is retained. When the morphology is broken, the skeleton point movement direction is reversed and attenuated. In this invention, when performing reverse attenuation, the spectral-morphological consistency offset gate reads the skeleton point movement direction, reverses the skeleton point movement direction to obtain the reverse movement direction; reads the original offset of the skeleton point, multiplies the original offset by the reverse attenuation coefficient to obtain the reverse attenuation offset; the skeleton point performs position update along the reverse movement direction according to the reverse attenuation offset, and the reverse attenuation coefficient is greater than 0 and less than 1. The inland points processed by the main response gating branch, the contour points processed by the boundary continuous gating branch, and the skeleton points processed by the direction extension gating branch are output according to the particle number binding relationship to obtain the inland points, contour points, and skeleton points after spectral-morphological consistency offset gate correction.

[0039] In this invention, the spectral-morphological consistency offset gate retains the particle numbers corresponding to the interior points, contour points, and skeleton points when receiving the point set output of the cascaded heterogeneous point set layer. The main body response gated branch retains the particle numbers after processing the interior points, the boundary continuity gated branch retains the particle numbers after processing the contour points, and the directional extension gated branch retains the particle numbers after processing the skeleton points. After the three gated branches are processed, the spectral-morphological consistency offset gate outputs the interior points, contour points, and skeleton points corresponding to the same particle according to the particle number binding relationship, forming the interior points, contour points, and skeleton points corrected by the spectral-morphological consistency offset gate.

[0040] In this embodiment, the processes of microplastic discrimination, particle counting, particle size determination, and fiber morphology differentiation in S6 specifically include: The particle body is determined by the inner domain points after spectral-morphological consistency offset gate correction. When the inner domain points are located in the stable region of the particle body and the contour points and skeleton points correspond to the inner domain points according to the particle number binding relationship, the corresponding particles are identified as microplastic particles. The particle body identification process in this invention includes: extracting the inner domain points corresponding to the same particle according to the particle number binding relationship after spectral-morphological consistency offset gate correction; counting whether the inner domain points corresponding to the same particle are located within the particle body stable region; when the number of inner domain points located within the particle body stable region reaches a set threshold for the number of main body points, the region corresponding to the particle is regarded as the particle body; inner domain points located outside the particle body stable region do not participate in the particle body identification; the set threshold for the number of main body points is determined according to the number of inner domain points in the cascaded heterogeneous point set layer, and is taken as 30% to 50% of the number of inner domain points in the candidate point set of the same particle; The execution process based on the particle number binding relationship and the correspondence between the inner domain points specifically includes: when the spectral-morphological consistency offset gate outputs the corrected inner domain points, contour points, and skeleton points, it retains the particle numbers assigned by the cascaded heterogeneous point set layer; it reads the inner domain points, contour points, and skeleton points according to the same particle number; when the particle number carried by the contour point and skeleton point is consistent with the particle number carried by the inner domain point, it is determined that the contour point, skeleton point, and inner domain point correspond; points with inconsistent particle numbers do not participate in the microplastic discrimination, particle counting, particle size determination, and fiber morphology differentiation of that particle. The process of identifying particles as microplastic particles specifically includes: reading the inner domain points, contour points, and skeleton points corresponding to the same particle according to the particle number binding relationship; when the inner domain point is located within the stable area of ​​the particle body, the contour point is distributed near the outer edge of the particle body, and the skeleton point and the contour point correspond to the same particle number, the particle corresponding to that particle number is identified as a microplastic particle; when the inner domain point deviates from the stable area of ​​the particle body, the contour point does not form the outer edge of the particle body, and the skeleton point and the inner domain point do not belong to the same particle number, the particle corresponding to that particle number is not counted as a microplastic particle. Determine the inner domain point, outline point, and skeleton point corresponding to each microplastic particle according to the particle number binding relationship, and accumulate the number of microplastic particles according to each particle number to complete the particle counting. In this invention, the particle numbers identified as microplastic particles are deduplicated; the inner field points, contour points, and skeleton points under the same particle number are counted as only one microplastic particle; when there are several inner field points, several contour points, and several skeleton points under the same particle number, they are not counted by the number of points, but by the particle number; the particle count is obtained by accumulating the total number of particle numbers identified as microplastic particles. Based on the particle number binding relationship, select the contour points corresponding to the same particle, take the distance between the two contour points with the largest distance as the major axis scale, and take the maximum span of the contour points perpendicular to the major axis scale as the minor axis scale. In this invention, contour points corresponding to the same particle are selected according to the particle number binding relationship; among the contour points corresponding to the same particle, the spatial distance between any two contour points is calculated one by one; the two contour points with the largest spatial distance are selected as the endpoints of the major axis; the spatial distance between the two endpoints of the major axis is used as the major axis scale; the endpoints of the major axis and the major axis scale are bound and saved with the particle number. The direction of the line connecting the endpoints of the major axis is taken as the major axis scale direction; among the contour points corresponding to the same particle, the vertical projection distance from each contour point to the major axis scale direction is calculated; the contour point on the side with the largest vertical projection distance and the contour point on the opposite side are selected; the maximum span between the two contour points in the direction perpendicular to the major axis scale is taken as the minor axis scale; the minor axis scale is bound to the particle number and saved. According to the particle number binding relationship, select the skeleton point, major axis scale and minor axis scale corresponding to the same particle. When the skeleton point is continuously distributed along the center line of the slender particle and the ratio of the major axis scale to the minor axis scale reaches the set aspect ratio threshold, the corresponding microplastic particles are classified into fiber morphology. The continuous distribution determination process in this invention includes: selecting skeleton points corresponding to the same particle according to the particle number binding relationship; projecting the skeleton points onto the direction of the center line of the slender particle; sorting the skeleton points according to the projection position; calculating the projection spacing and directional angle between adjacent skeleton points; when the projection spacing between adjacent skeleton points does not exceed the set skeleton spacing threshold and the directional angle between adjacent skeleton points does not exceed the set directional angle threshold, it is determined that the skeleton points are continuously distributed along the direction of the center line of the slender particle; the set skeleton spacing threshold is determined based on the average spacing of skeleton points along the direction of the center line of the slender particle, and is taken as 1.2 to 2 times the average spacing of skeleton points; the set directional angle threshold is taken as 10 degrees to 20 degrees. Based on the particle number binding relationship, the skeleton points, major axis scale, and minor axis scale corresponding to the same particle are selected; the aspect ratio threshold is set according to the slender morphological characteristics of fibrous microplastics, with a value of 3 to 5; when the skeleton points are continuously distributed along the center line of the slender particle, and the ratio of the major axis scale to the minor axis scale reaches the set aspect ratio threshold, the microplastic particle corresponding to that particle number is classified as fibrous morphology; the fibrous morphology result is bound to the particle number and used for subsequent particle size-concentration inversion output head calls; According to the particle number binding relationship, select the skeleton point corresponding to the same particle. When the skeleton point is not continuously distributed along the center line of the slender particle, the corresponding microplastic particle is classified as non-fiber particle. In this invention, the skeleton points corresponding to the same particle are selected according to the particle number binding relationship. When the skeleton points are not continuously distributed along the center line of the slender particle, the microplastic particles corresponding to that particle number are classified as non-fibrous particles. The major axis scale and minor axis scale corresponding to the same particle are selected according to the particle number binding relationship. When the ratio of the major axis scale to the minor axis scale does not reach the set aspect ratio threshold, the microplastic particles corresponding to that particle number are classified as non-fibrous particles. The non-fibrous particle results are bound to the particle number and used for subsequent particle size-concentration inversion output head calls. Based on the particle number binding relationship, the major axis and minor axis dimensions corresponding to the same particle are selected. When the ratio of the major axis dimension to the minor axis dimension does not reach the set aspect ratio threshold, the corresponding microplastic particle is classified as a non-fiber particle.

[0041] In this embodiment, the processing of the particle size-concentration inversion output head in S7 specifically includes: Receive the interior points, contour points, and skeleton points after spectral-morphological consistency offset gate correction, call the sampling volume, and call the major axis scale, minor axis scale, fiber morphology and non-fiber particles corresponding to each microplastic particle according to the particle number binding relationship. For microplastic particles classified into fibrous morphology, the distance between adjacent skeleton points is accumulated along the direction of skeleton point movement, and the cumulative result is used to correct the major axis scale. The corrected major axis scale is then used as the particle scale. In this invention, for microplastic particles classified into fibrous forms, the particle size-concentration inversion output head reads the skeleton points corresponding to the same particle according to the particle number binding relationship; the skeleton points are projected onto the skeleton point movement direction, and the skeleton points are arranged from small to large according to the projection position; the distance between adjacent skeleton points is calculated, and the distances of each adjacent point are added together to obtain the skeleton extension distance used to correct the major axis scale; when the skeleton extension distance is greater than the major axis scale, the major axis scale is replaced by the skeleton extension distance; when the skeleton extension distance is not greater than the major axis scale, the major axis scale is retained. For microplastic particles that are classified as non-fibrous particles, the particle size is determined using the long axis and short axis dimensions. In this invention, for microplastic particles classified as non-fibrous particles, the particle size-concentration inversion output head reads the major axis scale and minor axis scale corresponding to the same particle according to the particle number binding relationship; the average of the major axis scale and minor axis scale is calculated to obtain the particle scale of the non-fibrous particles; this particle scale is bound to the corresponding particle number and used for subsequent particle size range classification. For microplastic particles that retain the regression direction through the main response gated branch and whose skeleton points under the same particle number do not undergo reverse attenuation through the directional extension gated branch, the full count weight is assigned; for microplastic particles that undergo reverse attenuation through the directional extension gated branch, the reduced count weight is assigned. In this invention, the particle size-concentration inversion output head reads the processing results of the main response gating branch on the inner domain points; when the inner domain points under the same particle number retain the regression direction through the main response gating branch, the microplastic particles corresponding to the particle number are assigned a complete count weight; the complete count weight is set to 1, which is used to indicate that the particle participates in the concentration inversion as a complete and effective particle. The particle size-concentration inversion output head reads the processing results of the directional extension gating branch on the skeleton points; when a skeleton point under the same particle number undergoes reverse attenuation through the directional extension gating branch, the microplastic particle corresponding to that particle number is assigned a reduced count weight; the reduced count weight is obtained by multiplying the full count weight by the attenuation ratio, where the attenuation ratio is greater than 0 and less than 1. The counting weight of each microplastic particle is assigned to the corresponding particle size interval according to the particle size. The counting weights in each particle size interval are summed to obtain the weighted number of particles in each particle size interval. In this invention, the particle size-concentration inversion output head presets several particle size intervals; for each microplastic particle, the particle size bound to the particle number is read; the particle size is compared with the upper and lower limits of the particle size interval; when the particle size falls into a certain particle size interval, the count weight of the microplastic particle is written into the particle size interval; the count weight of the same particle number is only written into one particle size interval. Within each particle size range, the particle size-concentration inversion output head sums the count weights of the microplastic particles falling into that range to obtain the weighted particle count for that range; particles with full count weights are counted as 1, and particles with reduced count weights are counted according to the attenuation ratio. The weighted number of particles in each particle size range is divided by the sampling volume to obtain the microplastic concentration in each particle size range. The microplastic concentrations in each particle size range are summed to obtain the microplastic concentration detection results in the water body.

[0042] Example 1: To verify the feasibility of this invention in practice, it was applied to a comprehensive water sample microplastic detection scenario. The source type of the water sample to be tested was not specifically limited; the sample contained silt, organic debris, bubbles, non-plastic fibers, and microplastic particles of different morphologies. Microplastic particles included fragmented microplastics, sheet-like microplastics, and fibrous microplastics. Conventional image detection methods in such water samples easily identify the edges of silt, organic debris, and bubbles as the outer edges of microplastics. Fiber-like microplastics, affected by bending, breakage, uneven fluorescence response, and discontinuous boundaries, are easily broken into several small particles, leading to an overestimation of particle count, statistical deviation in particle size, and errors in concentration conversion.

[0043] This embodiment selected 120 water samples. To verify the applicability under different water sample conditions, the water samples were divided into low-turbidity water samples, high-impurity water samples, water samples with a high proportion of fibrous microplastics, and water samples with a high proportion of fragmented microplastics. All groups of water samples were subjected to the same imaging conditions, the same sampling volume recording method, and the same manual verification standard. A total of 1846 microplastic particles were identified through manual verification, and the results of the manual verification served as the benchmark for comparison of identification accuracy, particle size error, and concentration error. Conventional image detection methods use image segmentation and particle classification processes as a reference; the method of this invention uses a spectral-morphological coupled RepPoints network to complete the detection.

[0044] During the detection process, after water samples are enriched and processed, images of the water samples are acquired, and the corresponding sampling volume is recorded. Once the water sample images are fed into the spectral-morphological coupled RepPoints network, the spectral response channel reads the response difference between the particle region and the background region, while the particle morphology texture channel reads the particle edges, texture direction, and skeleton trend. The cross-coding layer incorporates spectral response information into the particle morphology texture encoding process and vice versa, ensuring that abnormal edges formed by silt, organic debris, and bubbles no longer independently dominate the particle recognition results.

[0045] After cross-coding, the RepPoints point set representation layer was transformed into a cascaded heterogeneous point set layer. This layer divides ordinary representative points into interior points, contour points, and skeleton points. Interior points first converge towards the stable region of the particle body, and the retained interior points form the convergence range of the interior points. Contour points use the convergence range of the interior points as the offset boundary and perform edge-fitting regression along the outer edge of the particle body. When a contour point crosses the outer edge of the offset boundary, it is projected to the nearest position on the offset boundary and then pressed back along the direction pointing towards the outer edge of the particle body. Skeleton points move along the centerline of the slender particles; the direction of skeleton point movement is written into the contour point edge-fitting regression process to correct the arrangement of contour points on both sides of the fibrous microplastic.

[0046] The interior points, contour points, and skeleton points output from the cascaded heterogeneous point set layer enter the spectral-morphological consistency offset gate. Interior points enter the main body response gating branch, contour points enter the boundary continuity gating branch, and skeleton points enter the direction extension gating branch. Points with consistent spectral responses and continuous morphology retain their regression direction, while points with discrete spectral responses and fragmented morphology undergo offset amplitude reduction or reverse attenuation. The corrected interior points are used to confirm the particle body and complete particle counting, the corrected contour points are used to determine the major and minor axis scales, and the corrected skeleton points are used to determine fiber morphology. This process prevents impurity edges from being continuously adsorbed by contour points, and the fracture boundaries of fibrous microplastics can be reconstrained according to the movement direction of the skeleton points.

[0047] The particle size-concentration inversion output head receives the interior points, contour points, and skeleton points after spectral-morphology consistency offset gate correction. For fibrous microplastics, the distance between adjacent skeleton points is accumulated along the direction of skeleton point movement, and the accumulated result is used to correct the major axis scale; for non-fibrous particles, the particle scale is determined using the major axis and minor axis scales. Each microplastic particle enters the corresponding particle size interval according to the particle number binding relationship. The counting weight is accumulated within the particle size interval, and the weighted particle number of each particle size interval is divided by the sampling volume to obtain the microplastic concentration of each particle size interval. The microplastic concentration of each particle size interval is summed to obtain the water microplastic concentration detection result.

[0048] Table 1: Comparison of Rapid Detection Effects of Microplastics in Water under Different Water Sample Conditions

[0049] Table 1 shows that the method of the present invention outperforms conventional image detection methods in terms of overall recognition accuracy, group recognition accuracy, impurity false detection rate, particle size error, concentration error, and processing time. The overall recognition accuracy increased from 85.2% to 94.7%, indicating that the cascaded heterogeneous point set layer, by separately representing the particle body, particle outer edge, and the centerline of slender particles, provides higher positioning stability for irregular microplastic particles. The recognition accuracy for high-impurity water samples increased from 79.6% to 91.2%, a greater improvement than for low-turbidity water samples, indicating that the convergence range of the inner domain points and the spectral-morphological consistency shift gate have a strong suppressive effect on interfering edges such as silt, organic debris, and bubbles.

[0050] The accuracy of fibrous microplastic identification improved from 81.5% to 93.1%, a result directly related to the process of embedding skeleton points into contour points for regression. Conventional image detection methods tend to segment the same fiber into several fragments when dealing with curved fibers and locally broken fibers. This invention corrects the arrangement of contour points on both sides by adjusting the direction of skeleton point movement, ensuring that the central trend and outer contour of the fibrous microplastic maintain a corresponding relationship, resulting in more stable fiber morphology differentiation.

[0051] The average particle size error in Table 1 decreased from 8.7 μm to 3.2 μm, and the relative error in concentration calculation decreased from 10.4% to 3.8%. This result indicates that after the corrected inner-area points, contour points, and skeleton points are directly connected to the particle size-concentration inversion output head, the connection between particle size and concentration calculation is closer. The average processing time for a single water sample was reduced from 126 seconds to 54 seconds, demonstrating that this invention places microplastic discrimination, particle counting, particle size determination, fiber morphology differentiation, and concentration conversion in a continuous processing chain, reducing the workload of manual correction and post-processing statistics.

[0052] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A rapid detection method for water microplastics based on multi-scale feature fusion, characterized in that, Includes the following steps: S1. Obtain the water sample image and corresponding sampling volume, and send the water sample image into the spectral-morphological coupling RepPoints network to perform cross-coding of the spectral response channel and the particle morphology texture channel. S2. Transform the RepPoints point set expression layer into a cascaded heterogeneous point set layer, and perform point set updates in the order of inner domain points, contour points, and skeleton points, so that the inner domain points first converge to the stable region of the particle body. S3. Limit the offset boundary of the contour point by the convergence range of the inner domain point, so that the contour point performs edge regression along the outer edge of the particle body, and restricts the contour point from expanding outward to the edge of the impurity. S4. Embed the skeleton points into the contour point regression process, so that the skeleton points extend along the center line of the slender particles and then the contour point arrangement is corrected in the reverse direction to complete the morphological constraint of the fibrous microplastics. S5. After the cascaded heterogeneous point set layer, a spectral-morphological consistency offset gate is set to retain the regression direction for the three types of points with consistent spectral response and continuous morphology, and to perform reverse attenuation for the three types of points with discrete spectral response and broken morphology. S6. Microplastic discrimination, particle counting, particle size determination and fiber morphology differentiation are performed using the inner domain points, contour points and skeleton points after spectral-morphological consistency offset gate correction. S7. Connect the inner domain points, contour points, and skeleton points after spectral-morphology consistency offset gate correction to the particle size-concentration inversion output head, and convert the water microplastic concentration detection results according to particle count, particle size, fiber morphology, and sampling volume.

2. The rapid detection method for water microplastics based on multi-scale feature fusion according to claim 1, characterized in that, The spectral-morphology coupled RepPoints network in S1 includes a spectral response channel, a particle morphology texture channel, a cross-coding layer, a cascaded heterogeneous point set layer modified from the RepPoints point set expression layer, a spectral-morphology consistency shift gate, and a particle size-concentration inversion output head. The spectral response channel and the particle morphology texture channel are connected to the cross-coding layer, the cross-coding layer is connected to the cascaded heterogeneous point set layer, the cascaded heterogeneous point set layer is connected to the spectral-morphology consistency shift gate, and the spectral-morphology consistency shift gate is connected to the particle size-concentration inversion output head.

3. The rapid detection method for water microplastics based on multi-scale feature fusion according to claim 2, characterized in that, The cross-coding in S1 specifically includes: By compressing adjacent spectral responses within a spectral response channel through one-dimensional channel convolution, spectral gate weights are formed. These spectral gate weights are then multiplied into the edge convolution path of the grain morphology texture channel to obtain the edge convolution encoding updated along the stable position of the spectral response. The boundary orientation and skeleton orientation within the particle morphology texture channel are extracted by directional convolution to form morphology gating weights. The morphology gating weights are multiplied into the channel convolution path of the spectral response channel to obtain the channel convolution encoding updated along the continuous position of the morphology. The edge convolutional encoding and channel convolutional encoding are added position by position; The code obtained by adding position by position is written back to the spectral response channel and the grain morphology texture channel through residual connection, respectively, to complete the cross-coding of the spectral response channel and the grain morphology texture channel.

4. The rapid detection method for water microplastics based on multi-scale feature fusion according to claim 3, characterized in that, The process of updating the point set in S2 specifically includes: Using the inner domain points as the first update points, the inner domain points are moved according to the direction from the low response position to the high response position in the encoding output of the cross-coding layer. The inner domain points that deviate from the stable region of the particle are removed, and the remaining inner domain points form the convergence range of the inner domain points. Use the contour points as the second-level update points to bind the convergence range of the contour points to the interior points; Use the skeleton point as the third-level update point to bind the skeleton point to the contour point corresponding to the same particle; Maintain the particle number binding relationship between the inland points, contour points, and skeleton points of the same particle, and output the inland points, contour points, and skeleton points as point sets of the cascaded heterogeneous point set layer.

5. The rapid detection method for water microplastics based on multi-scale feature fusion according to claim 4, characterized in that, The S3 section defines the contour point offset boundary by the convergence range of the inner domain points, specifically including: Using the convergence range of the inner domain points as the offset boundary of the contour points, the edge response direction of the encoding output of the cross-coding layer near the outer edge of the convergence range of the inner domain points is read. Perform edge-fitting regression on the contour points along the edge response direction toward the outer edge of the particle body; When the contour point crosses the outer edge of the offset boundary, project the contour point to the nearest position on the offset boundary and press the contour point back along the direction pointing to the outer edge of the particle body. When the contour point is located inside the offset boundary and fits the outer edge of the particle body, the contour point position is retained.

6. The rapid detection method for water microplastics based on multi-scale feature fusion according to claim 5, characterized in that, The process of embedding skeleton points into contour point regression in S4 specifically includes: The skeleton point reads the directional continuous response in the output of the cross-coding layer and moves along the center line of the slender particle. The direction of the line connecting adjacent skeleton points is used as the direction of skeleton point movement and written into the contour point edge-fitting regression process. Using the direction of skeleton point movement as an axial reference, the contour points on both sides of the slender particles are grouped and paired. Based on the distance difference between the paired contour points and the corresponding skeleton points, adjust the arrangement order and distance between the contour points on both sides to keep the contour points aligned with the shape of the slender particles.

7. The rapid detection method for water microplastics based on multi-scale feature fusion according to claim 6, characterized in that, The specific processing steps of the spectral-morphological consistency shift gate in S5 include: The inner domain points, contour points, and skeleton points in the output of the cascaded heterogeneous point set layer are input into the spectral-morphological consistency offset gate, so that the inner domain points enter the main response gated branch, the contour points enter the boundary continuity gated branch, and the skeleton points enter the directional extension gated branch. In the main response gating branch, the spectral response code is read along the regression direction of the inner domain point. When the spectral response is consistent, the regression direction of the inner domain point is retained. When the spectral response is discrete, the offset amplitude of the inner domain point is shortened. In the boundary continuous gating branch, the particle morphology texture code is read along the regression direction of the contour point. When the morphology is continuous, the regression direction of the contour point is retained. When the morphology is broken, the contour point is pressed back to the outer edge of the particle body. In the directional extension gated branch, the particle morphology texture code is read along the skeleton point movement direction. When the morphology is continuous, the skeleton point movement direction is retained. When the morphology is broken, the skeleton point movement direction is reversed and attenuated. The inland points processed by the main response gating branch, the contour points processed by the boundary continuous gating branch, and the skeleton points processed by the direction extension gating branch are output according to the particle number binding relationship to obtain the inland points, contour points, and skeleton points after spectral-morphological consistency offset gate correction.

8. The rapid detection method for water microplastics based on multi-scale feature fusion according to claim 7, characterized in that, The process of microplastic discrimination, particle counting, particle size determination, and fiber morphology differentiation in S6 specifically includes: The particle body is determined by the inner domain points after spectral-morphological consistency offset gate correction. When the inner domain points are located in the stable region of the particle body and the contour points and skeleton points correspond to the inner domain points according to the particle number binding relationship, the corresponding particles are identified as microplastic particles. Determine the inner domain point, outline point, and skeleton point corresponding to each microplastic particle according to the particle number binding relationship, and accumulate the number of microplastic particles according to each particle number to complete the particle counting. Based on the particle number binding relationship, select the contour points corresponding to the same particle, take the distance between the two contour points with the largest distance as the major axis scale, and take the maximum span of the contour points perpendicular to the major axis scale as the minor axis scale. According to the particle number binding relationship, select the skeleton point, major axis scale and minor axis scale corresponding to the same particle. When the skeleton point is continuously distributed along the center line of the slender particle and the ratio of the major axis scale to the minor axis scale reaches the set aspect ratio threshold, the corresponding microplastic particles are classified into fiber morphology. According to the particle number binding relationship, select the skeleton point corresponding to the same particle. When the skeleton point is not continuously distributed along the center line of the slender particle, the corresponding microplastic particle is classified as non-fiber particle. Based on the particle number binding relationship, the major axis and minor axis dimensions corresponding to the same particle are selected. When the ratio of the major axis dimension to the minor axis dimension does not reach the set aspect ratio threshold, the corresponding microplastic particle is classified as a non-fiber particle.

9. The rapid detection method for water microplastics based on multi-scale feature fusion according to claim 8, characterized in that, The processing steps of the particle size-concentration inversion output head in S7 specifically include: Receive the interior points, contour points, and skeleton points after spectral-morphological consistency offset gate correction, call the sampling volume, and call the major axis scale, minor axis scale, fiber morphology and non-fiber particles corresponding to each microplastic particle according to the particle number binding relationship. For microplastic particles classified into fibrous morphology, the distance between adjacent skeleton points is accumulated along the direction of skeleton point movement, and the cumulative result is used to correct the major axis scale. The corrected major axis scale is then used as the particle scale. For microplastic particles that are classified as non-fibrous particles, the particle size is determined using the long axis and short axis dimensions. For microplastic particles that retain the regression direction through the main response gated branch and whose skeleton points under the same particle number do not undergo reverse attenuation through the directional extension gated branch, the full count weight is assigned; for microplastic particles that undergo reverse attenuation through the directional extension gated branch, the reduced count weight is assigned. The counting weight of each microplastic particle is assigned to the corresponding particle size interval according to the particle size. The counting weights in each particle size interval are summed to obtain the weighted number of particles in each particle size interval. The weighted number of particles in each particle size range is divided by the sampling volume to obtain the microplastic concentration in each particle size range. The microplastic concentrations in each particle size range are summed to obtain the microplastic concentration detection results in the water body.

Images