Marine plankton image segmentation method and system

Abstract

The invention relates to the technical field of image data processing, in particular to an image segmentation method and system applied to marine plankton. The method comprises the following steps: acquiring an in-situ microscopic image of marine plankton, carrying out binarization processing and connected domain analysis on the in-situ microscopic image, and screening out a biological adhesion block mass to be processed; synchronously extracting the edge contour of the biological adhesion block mass to be processed, and calculating a biological structure integrity score; selecting the candidate segmentation path with the highest biological structure integrity score as an optimal segmentation line; curvature smooth completion is carried out along the optimal segmentation line, and a single biological image is obtained; and inputting the single biological image into a pre-trained classifier to obtain a biological abundance data set. According to the method, high-precision automatic identification and abundance statistics of plankton in a complex marine environment are realized by screening and positioning the adhesion block mass, carrying out adaptive adhesion segmentation and combining multi-dimensional morphological feature classification.

Description

Technical Field

[0001] This invention relates to the field of image data processing technology, and in particular to a method and system for image segmentation of marine planktonic organisms. Background Technology

[0002] In actual in-situ imaging, plankton are often in a three-dimensional swimming state, and in red tides or high-abundance waters, they form severe overlaps and adhesions. Simultaneously, different species exhibit significant morphological differences, with many plankton possessing slender tentacles, setae, or complex limb protrusions. These factors combine to result in blurred target edges and overlapping textures in the image, easily confusing the natural morphological boundaries of individual organisms with adhesion boundaries, ultimately affecting the accuracy of biodiversity statistics. Traditional watershed algorithms or segmentation strategies based on simple geometric concave points, commonly used in marine plankton image segmentation techniques, suffer from logical flaws due to their inability to distinguish between "biological limb connections" and "inter-individual adhesions." Furthermore, they easily misjudge complex protruding structures of the organism itself (such as the roots of tentacles and setae) as adhesion boundaries, leading to over-segmentation of individual organisms, generating numerous fragmented false targets, and causing severe distortion of the final biodiversity statistics. Summary of the Invention

[0003] Based on this, the present invention provides a method and system for image segmentation of marine plankton to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above objectives, a method for segmenting marine plankton images includes the following steps: Step S1: Obtain in-situ microscopic images of marine plankton, perform binarization processing and connected component analysis on the in-situ microscopic images, filter out target areas with an area exceeding the preset individual area threshold, and mark the target areas as biological aggregates to be processed. Step S2: Simultaneously extract the edge contour of the biological adhesion mass to be processed to locate the contour concavity extreme points and construct a topological structure map at the same time; construct multiple candidate segmentation paths based on the set of contour concavity extreme points, map the candidate segmentation paths to the topological structure map, and calculate the biological structure integrity score by combining the topological fracture cost of the candidate segmentation paths. Step S3: Select the candidate segmentation path with the highest biological structure integrity score as the optimal segmentation line; Step S4: Separate the biological adhesion clumps to be processed into independent sub-regions along the optimal segmentation line, and perform curvature smoothing to complete the separated cut edges to obtain individual biological images; Step S5: Input the individual biological image into the pre-trained classifier to obtain a biological abundance dataset containing species category, number of individuals, and volume concentration.

[0005] The present invention also provides a marine plankton image segmentation system that performs the marine plankton image segmentation method described above. The marine plankton image segmentation system includes: The agglomerate screening module is used to acquire in-situ microscopic images of marine plankton, perform binarization processing and connected component analysis on the in-situ microscopic images, screen out target areas with an area exceeding a preset individual area threshold, and mark the target areas as agglomerates of organisms to be processed. The structural integrity analysis module is used to simultaneously extract the edge contours of the biological aggregates to be processed, locate the contour depression extreme points, and construct a topological structure map. Based on the set of contour depression extreme points, multiple candidate segmentation paths are constructed, the candidate segmentation paths are mapped to the topological structure map, and the biological structural integrity score is calculated by combining the topological fracture cost of the candidate segmentation paths. The optimal segmentation decision module is used to select the candidate segmentation path with the highest biological structure integrity score as the optimal segmentation line. The single-organism separation and repair module is used to separate the biological adhesion clumps to be processed into independent sub-regions along the optimal segmentation line, and to perform curvature smoothing and completion on the separated cut edges to obtain single-organism images; The biological classification and statistics module is used to input individual biological images into a pre-trained classifier to obtain a biological abundance dataset that includes species category, number of individuals, and volume concentration.

[0006] The beneficial effects of this invention are as follows: By simultaneously extracting the edge contours and internal topological skeletons of the bio-adhesive clumps to be processed, and constructing candidate segmentation paths by combining concave extreme points, and introducing a node-type-based fracture impedance evaluation mechanism into the topological structure map, structured and accurate segmentation of complex bio-adhesive images is achieved. During segmentation decision-making, multi-dimensional morphological representation vectors such as contour frequency domain harmonic coefficients and regional geometric topological factors are used to assist classification, accurately distinguishing between the organism's own limb connections (high-impedance trunk) and the adhesion boundaries between individuals (low-impedance connections), thereby calculating the bio-structural integrity score of the candidate paths. This segmentation method based on topological semantic constraints effectively avoids the "over-segmentation" phenomenon caused by the traditional watershed algorithm's inability to identify the organism's intrinsic structure, and avoids misjudging tiny protrusions such as tentacles and setae as independent targets, providing a reliable morphological basis for the accurate counting of high-density plankton.

[0007] In the post-segmentation processing stage, independent sub-regions are separated along the optimal segmentation line, and the cut edges are smoothed and completed. A recursive residual detection mechanism based on sub-region area and convex hull filling rate is implemented to ensure the integrity and naturalness of individual biological morphologies. By iteratively judging whether the segmentation result still contains residual, adhered clumps and automatically triggering secondary segmentation, the problem of missed segments in severely stacked states can be completely solved. This closed-loop segmentation and repair strategy avoids statistical omissions caused by incomplete segmentation in a single step, effectively improving the confidence of the biological abundance dataset.

[0008] By normalizing and weighting the aspect ratio characteristics of the concavity in geometric space and the fracture impedance value in topological space, a biological structure integrity score is calculated. This data processing method efficiently integrates visual edge information with biological anatomical constraints into a decision index with clear physical meaning. In practical applications, whether dealing with high-density samples during red tide outbreaks or multi-pose swimming targets in deep-sea in-situ exploration, it can quickly select the segmentation scheme that best conforms to the natural growth patterns of organisms and transform it into accurate species classification statistics. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the steps in the method for segmenting marine planktonic images according to the present invention; Figure 2 This is a schematic diagram of the screening of marine planktonic agglomerates in this invention; Figure 3 This is a schematic diagram illustrating the segmentation and cut completion of planktonic aggregates in this invention; Figure 4 This is a schematic diagram of the marine planktonic image segmentation process of the present invention; The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0010] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0011] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0012] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0013] To achieve the above objectives, please refer to Figures 1 to 4 This invention provides a method for image segmentation of marine plankton, comprising the following steps: Preferably, step S1: acquire in-situ microscopic images of marine plankton, perform binarization processing and connected component analysis on the in-situ microscopic images, screen out target areas with an area exceeding a preset individual area threshold, and mark the target areas as biological aggregates to be processed. In one embodiment, an in-situ microscopic imager mounted on an underwater vehicle is used to take pictures at a predetermined water depth. The imager is equipped with a high-frequency flash to eliminate motion blur and acquire in-situ microscopic images with a resolution of not less than 2048×2048 pixels.

[0014] In one implementation of this invention, binarization of the in-situ microscopic image specifically includes: first, converting the color in-situ microscopic image into a grayscale image; then, calculating the grayscale value of each pixel using a weighted average method, with the grayscale value calculated as: grayscale value = 0.299 × red component + 0.587 × green component + 0.114 × blue component; subsequently, calculating the optimal segmentation threshold using the maximum inter-class variance method; traversing each pixel in the image; marking pixels with grayscale values ​​greater than the optimal segmentation threshold as foreground target points and assigning them a value of 255; and marking pixels with grayscale values ​​less than or equal to the optimal segmentation threshold as background points and assigning them a value of 0, thereby generating a binarized mask image.

[0015] Specifically, the operation process for connected component analysis of the binarized image is as follows: the 8-neighborhood connectivity algorithm is used to scan the binarized mask image. When a foreground pixel with a pixel value of 255 is detected, the pixel status of its 8 neighboring positions is checked. If there is a labeled connected region in the neighboring positions, the current pixel is assigned to that connected region; otherwise, a new region label is assigned. After the scan is completed, the total number of pixels contained in each independent connected region is counted, and the total number of pixels is defined as the area value of the region.

[0016] In one operation step of this embodiment, the specific steps for filtering out target regions whose area exceeds the preset single-unit area threshold are as follows: read the area value of each region obtained from the current connected component analysis, compare it one by one with the preset single-unit area threshold, and at the same time crop the region image of the bio-adhesive clump to be processed and save it to a temporary processing queue.

[0017] Please see Figure 2 This is a schematic diagram illustrating the screening of marine planktonic agglomerates in this invention. The diagram shows the agglomerate region under in-situ microscopic imaging: the main part of the image consists of agglomerates formed by multiple plankton (such as dinoflagellates with a dark central nucleus and radial patterns on the periphery) adhering to each other. The range of the agglomerates to be processed is marked by black outlines. The small circular structures within the agglomerates (marked with black boxes) are small agglomerated plankton individuals. Target areas exceeding the individual threshold have been screened using binarization and connected component analysis, clarifying the boundaries of the agglomerates to be segmented subsequently. The background is the grayscale base of the microscopic image, and individual plankton individuals scattered at the edges of the image (such as those with pores in the upper left corner) are visible as non-agglomerated reference samples.

[0018] Preferably, step S2: Simultaneously extract the edge contour of the biological adhesion mass to be processed to locate the contour concavity extreme points, and at the same time construct a topological structure map; construct multiple candidate segmentation paths based on the set of contour concavity extreme points, map the candidate segmentation paths to the topological structure map, and calculate the biological structure integrity score by combining the topological fracture cost of the candidate segmentation paths; Optionally, step S2, which involves simultaneously extracting the edge contour of the bio-adhesive aggregates to be processed in order to locate the extreme points of the contour depressions, includes: The covariance matrix of all pixel coordinates within the window is calculated by moving the preset sliding sampling window along the edge contour of the biological adhesion mass to be processed. The discrete curvature value of the center pixel of the window is determined based on the ratio of eigenvalues ​​of the covariance matrix, and the curvature distribution curve of the whole contour is generated. The extreme value search is performed on the curvature distribution curve of the whole contour to identify the trough position where the curvature value is lower than the adjacent pixel point, and it is defined as the local curvature minimum point. Set a negative curvature threshold, remove fluctuations with curvature values ​​greater than the negative curvature threshold from the local curvature minima, and determine the remaining local curvature minima as contour concavity extreme points.

[0019] In one embodiment, the Canny edge detection operator is used to extract the closed edge contour of the biological adhesion mass to be processed by a single pixel width, and the pixels on the edge contour are sorted in a clockwise direction to form an ordered sequence of edge pixel coordinates; the size of the preset sliding sampling window is usually set to an odd number.

[0020] In one implementation of this embodiment, the specific operation for calculating the covariance matrix of all pixel coordinates within the window is as follows: First, calculate the average abscissa and ordinate of all pixels in the local sampling set, and use them as the local centroid coordinates; then, traverse each pixel within the window, calculate the deviation vector between the pixel coordinates and the local centroid coordinates, multiply the deviation vector by its transpose to obtain a 2×2 matrix, add the matrices corresponding to all pixels within the window, and divide by the total number of pixels within the window minus 1, thereby obtaining the covariance matrix describing the directionality and dispersion of pixel distribution in the local region.

[0021] Specifically, determining the discrete curvature value of the window center pixel based on the eigenvalue ratio of the covariance matrix includes: performing eigenvalue decomposition on the aforementioned 2×2 covariance matrix to obtain larger eigenvalues. and smaller eigenvalues Calculate the ratio of the two. As a measure of anisotropy, a first vector is constructed pointing from the start point of the window to the center point, and a second vector is constructed pointing from the center point to the end point of the window. The cross product of the first and second vectors is calculated. If the cross product is negative, it is determined to be a concave region, and a ratio is assigned. A negative sign indicates a convex region if the cross product is positive, and a ratio is assigned accordingly. The positive sign is used to obtain the signed discrete curvature values, and the full contour curvature distribution curve is generated in the order of the pixel sequence.

[0022] It should be noted that the method for setting the negative curvature threshold is as follows: Select no less than 50 planktonic organism samples with smooth surfaces and no adhesion, calculate the discrete curvature value of their edge contours according to the above steps, and count the minimum negative curvature value (i.e. the negative value with the largest absolute value) that appears in all samples. For example, the minimum negative curvature value obtained by statistics is -0.15. In order to avoid misjudgment caused by image digitization noise, multiply the statistical value by a safety factor of 1.2, and finally determine the preset negative curvature threshold as -0.18.

[0023] Optionally, the construction of the topology diagram in step S2 includes: The biological adhesion clumps to be treated are peeled off layer by layer until a central skeleton line with a single pixel width is retained. Calculate the eight-neighbor connectivity of each pixel on the central skeleton line, mark points with a connectivity greater than two as skeleton branch points, and mark points with a connectivity of one as skeleton endpoints; Tracing the connection paths between skeleton branch points, a weighted undirected graph is constructed with skeleton branch points and skeleton endpoints as nodes and connection paths as edges. The pixel length of the connection path is used as the edge weight of the weighted undirected graph to form a topological structure graph.

[0024] In one embodiment, the layer-by-layer peeling operation is implemented using a morphological thinning algorithm. The binary image of the biological adhesion mass to be processed is used as input, and a 3×3 structuring element is defined. The edge pixels of the image are scanned cyclically to determine whether the current foreground pixel meets the deletion conditions. The deletion conditions include that the point is not an endpoint and that deletion does not destroy local connectivity. The above edge peeling operation is repeated until all foreground regions converge to a central skeleton line with a width of one pixel. At this time, the horizontal or vertical width of any point on the skeleton line is 1 pixel.

[0025] In one embodiment, calculating the eight-neighbor connectivity of a pixel on the central skeleton line specifically involves: traversing each pixel on the central skeleton line and constructing a 3×3 neighborhood window centered on that pixel; counting the number of foreground pixels in the eight adjacent positions (excluding the center point) within the neighborhood window, which is the eight-neighbor connectivity. For example, if a pixel has only one foreground pixel in its eight-neighbor area, it is determined to be a skeleton endpoint, representing the start and end positions of the skeleton; if a pixel has three or more foreground pixels in its eight-neighbor area, it is determined to be a skeleton branch point, representing the intersection or fork position of the skeleton; if the connectivity is 2, it is considered a normal skeleton path point.

[0026] Specifically, the connection paths between skeleton branch points are traced and a topological graph is constructed. Starting from any marked skeleton branch point, the search and tracing proceeds along the skeleton path points in adjacent directions until another skeleton branch point or skeleton endpoint is encountered. This continuous set of pixels is defined as a connection path. A weighted undirected graph is constructed with skeleton branch points and skeleton endpoints as nodes and connection paths as edges. The calculation method for the pixel length of the connection path as the edge weight of the weighted undirected graph is as follows: Initialize the cumulative length of the current connection path to 0; during path tracing, determine the displacement direction after passing each adjacent pixel point; if the movement is horizontal or vertical, the cumulative length value increases by 1; if the movement is diagonal, the cumulative length value increases by 1.414; or simply count the total number of pixels contained in the path as the length value.

[0027] Of particular importance, step S2, which involves constructing a topology graph based on the internal morphological skeleton, including skeleton branch points, skeleton endpoints, and connection path weights, also includes: Perform Euclidean distance transformation on the binarized image of the biological adhesion mass to be processed to generate a distance transformation map, where the gray value of each pixel in the distance transformation map represents the Euclidean distance from that point to the nearest background pixel; The distance transformation map is iteratively refined to extract the central skeleton line with a single pixel width. The gray value of each pixel on the central skeleton line corresponding to the distance transformation map is assigned to that pixel as the width attribute value of the skeleton node. Traverse the central skeleton line, identify the original branch point, and calculate the number of skeleton branches in the preset neighborhood centered on the original branch point; if the number of skeleton branches is greater than the preset threshold, and the width attribute value of the skeleton node of the original branch point is less than the preset edge noise width threshold, then the original branch point is determined to be a pseudo branch point and is removed, and the remaining points are retained as valid skeleton branch points. A topology graph is constructed based on effective skeleton branch points and skeleton endpoints. The mean value of the skeleton node width attribute of all pixels on the connection path is defined as the radial thickness weight of the connection path and stored in the topology graph.

[0028] In one embodiment, performing Euclidean distance transformation on the binarized image of the biological adhesion mass to be processed specifically involves: traversing each foreground pixel in the binarized image, searching for the nearest background pixel, calculating the Euclidean geometric distance between the two, and linearly mapping the distance value to a grayscale value to generate a distance transformation map.

[0029] In one embodiment, the operation process for iteratively refining the distance transformation map is as follows: a connectivity-preserving morphological refining algorithm is used to peel off the foreground region of the distance transformation map layer by layer until the foreground region converges to a central skeleton line with a width of one pixel; then, each pixel on the central skeleton line is traversed, the gray value corresponding to the coordinates of the pixel in the original distance transformation map is read, and the gray value is directly assigned to the skeleton pixel as the width attribute value of the skeleton node. This attribute value physically represents the local radial radius of the organism at the skeleton point.

[0030] Specifically, the steps for identifying and removing pseudo-branch points include: identifying all original branch points with an eight-neighbor connectivity greater than 2, establishing a 5×5 pixel preset neighborhood centered on each original branch point; counting the number of skeleton branches existing in the preset neighborhood; if the number of branches is greater than 3, and the width attribute value of the skeleton node of the original branch point itself is less than the preset edge noise width threshold calculated above, then the original branch point is determined to be a pseudo-branch point caused by edge noise and is removed, retaining only valid skeleton branch points that meet the conditions.

[0031] Optionally, in step S2, constructing multiple candidate segmentation paths based on the set of extreme points of the contour concavity includes: Calculate the Euclidean distance between all contour concave extreme points, and pair non-adjacent contour concave extreme points with a distance less than the preset adhesion width to form candidate segmentation point pairs; Connect candidate split point pairs to form a straight line segment, or extend along the normal direction of the two point pairs to intersect and form a broken line segment. Define the straight line segment or broken line segment as a candidate split path.

[0032] It should be noted that the preset adhesion width is obtained as follows: 200 typical biological adhesion images are randomly selected from the historical database, and the maximum width of the adhesion junction is manually measured; the average value of all measured widths is calculated and set as the baseline width, and 1.5 times the baseline width is used as the preset adhesion width. This threshold is designed to ensure that it can cover the width of the vast majority of adhesion necks.

[0033] Specifically, the operations for constructing candidate segmentation paths for each set of candidate segmentation points include: the first method is to construct straight line segments by directly using a line drawing algorithm to generate a set of straight pixel lines connecting the coordinates of two extreme points; the second method is to construct polyline segments by calculating the slope of the tangent lines at the contours of the two extreme points, thereby deriving the normal direction vector perpendicular to the tangent lines, extending rays along the normal direction into the interior of the biological cluster, and solving for the coordinates of the intersection point of the two rays; if the intersection point is located inside the biological cluster, then a polyline path is constructed connecting the first extreme point to the intersection point, and then connecting the intersection point to the second extreme point. Finally, all the generated straight line segments and polyline segments are uniformly defined as candidate segmentation paths and added to the processing list.

[0034] Optionally, before defining line segments or polyline segments as candidate segmentation paths, a validation of the pairing validity is also included: For the two contour concavity extreme points in the candidate segmentation point pair, search for the pixel point with the closest Euclidean distance on the central skeleton line of the topology map, and use it as the skeleton projection point. Search for the shortest skeleton path between two skeleton projection points in the topology graph and calculate the topological distance of the shortest skeleton path. Calculate the Euclidean distance between two extreme points of the contour concavity. If the ratio of the Euclidean distance to the topological distance is less than the preset path curvature threshold, the candidate segmentation point pair is determined to be valid and candidate segmentation paths are allowed to be generated; otherwise, it is determined to be an invalid pair.

[0035] Specifically, searching for the shortest skeleton path and calculating the topological distance in the topological graph includes: using Dijkstra's shortest path algorithm or breadth-first search algorithm to search for a path connecting the first skeleton projection point and the second skeleton projection point in the weighted undirected graph; during the search process, the weight values ​​of all edges traversed by the path are accumulated. Since the edge weights represent the pixel length of the connecting path, the final accumulated total weight value is the actual physical length of the shortest skeleton path connecting the two projection points, and it is recorded as the topological distance.

[0036] It should be noted that the preset path curvature threshold is obtained as follows: A test dataset containing both adhered and non-adhesive curved biological samples is established; the ratio of Euclidean distance to topological distance for each pair of potential segmentation points in the dataset is calculated; statistical analysis shows that this ratio is usually close to 1 for the curved parts of non-adhesive organisms, while the ratio is significantly smaller for the actual adhered parts due to their short spatial distance but long skeletal extension; a critical value that distinguishes the two types of samples is selected, for example, 0.6 is selected as the dividing point, and it is set as the preset path curvature threshold.

[0037] In one implementation of this invention, the calculation process for performing pairing validity verification is as follows: the length of the direct line connecting two extreme points of the contour concavity is read as the Euclidean distance, and the length of the shortest skeleton path obtained above is read as the topological distance; the Euclidean distance is divided by the topological distance to obtain the curvature ratio.

[0038] Optionally, in step S2, mapping the candidate segmentation paths to the topology graph and calculating the biological structure integrity score based on the topological fracture cost of the candidate segmentation paths includes: The depth-to-width ratio feature of the concave is calculated based on the extreme points of the concave contours at both ends of the candidate segmentation path. Detect whether the candidate segmentation path intersects with the connection path in the topology graph, and mark the skeleton edge at each intersection position as a topological interference edge; Assign topological interference coefficients based on the node types at both ends of the topological interference edge; The original edge weights of the topological interference edges are obtained from the topological structure graph, and the biological structure integrity score of the candidate segmentation path is calculated based on the indentation aspect ratio, the original edge weights and the topological interference coefficient.

[0039] In one embodiment of the present invention, based on the position of the contour concavity extreme points corresponding to both ends of the candidate segmentation path on the edge contour, the vertical distance from the contour concavity extreme point to the convex hull boundary is calculated as the concavity depth, and combined with the Euclidean distance between both ends of the candidate segmentation path, the depth-to-width ratio feature of the concavity is calculated.

[0040] In one embodiment, the specific steps for constructing the convex hull and determining the boundary line segments are as follows: obtain the coordinates of all pixel points on the edge contour of the biological adhesion mass to be processed, generate the smallest convex polygon surrounding these contour points using the Graham scan algorithm or the Jarvis step algorithm, and the boundary of the convex polygon is the convex hull boundary; decompose the convex hull boundary into several continuous line segments, each line segment being determined by two adjacent vertices on the convex hull.

[0041] In one embodiment of the present invention, a topological interference coefficient is assigned according to the node type at both ends of the topological interference edge; wherein, if both ends are skeleton branch points, it is determined that the torso is damaged and a first high interference coefficient is assigned; if one end is a skeleton endpoint, it is determined that the limb is severed and a second low interference coefficient is assigned.

[0042] In one embodiment, the specific operation for determining the topological interference edge is as follows: the currently generated candidate segmentation path is mapped to the coordinate system where the topological structure graph is located, and the line segment intersection detection algorithm is used to determine whether the path spatially intersects with any connecting path in the topological structure graph; if there is an intersection, the intersected connecting path is marked as a topological interference edge, and the node attributes connected at both ends of the edge are read from the data structure.

[0043] It should be noted that the numerical setting logic for the first high-value interference coefficient and the second low-value interference coefficient is as follows: Based on a large number of correctly segmented biological image samples, reverse training is performed to statistically determine the negative impact of mis-segmentation of the torso on recognition accuracy, as well as the necessity of amputating limbs (such as crossed tentacles). According to the statistical results, the first high-value interference coefficient is usually set to a value between 1.0 and 5.0 to ensure that the segmentation is suppressed with a high probability. The second low-value interference coefficient is usually set to a value between 0.1 and 0.3, which only produces a slight cost and allows segmentation to be performed under conditions where other features (such as concavity salience) are strong.

[0044] Specifically, the procedure for calculating the biological structure integrity score of candidate segmentation paths is as follows: The indentation aspect ratio feature, original edge weights, and topological interference coefficients are substituted into a pre-defined scoring model for weighted calculation; the mathematical expression of the scoring model is: ; in, Represents the score of biological structural integrity. Represents the aspect ratio of the depression. Represents the original edge weights. Represents the topological interference coefficients. and The preset balance factor weights.

[0045] Preferably, step S3: select the candidate segmentation path with the highest biological structure integrity score as the optimal segmentation line; Preferably, step S4: the biological adhesion clumps to be processed are separated into independent sub-regions along the optimal segmentation line, and the edges of the separated cuts are smoothed by curvature to obtain individual biological images; Please see Figure 3 This diagram illustrates the segmentation and incision repair of planktonic aggregates in this invention, showing four typical examples of segmentation and repair of planktonic aggregates: Example in the upper left corner (dinoflagellates): Red anchor points mark the intersection of the optimal dividing line and the clumping outline. The green curve is a cubic Bezier smooth completion curve generated based on the anchor point tangent slope constraint, which replaces the original straight dividing line and realizes the edge closure of the adhered individuals. Example in the upper right corner (diatom plankton): The blue dashed line is the original optimal dividing line, the green curve is the smooth edge after completion, the red anchor point corresponds to the intersection of the dividing line and the outline, and the arrow indicates the transition direction of the completed curve; Example in the lower left corner (foraminifera plankton): The yellow box marks the endpoints of the segmented cut, the blue dashed line is the original segmentation path, and the green curve is the virtual predicted boundary curve generated based on the contour curvature change rate, which completes the natural closure of the spiral edge; Example in the lower right corner (radiolarvae): The yellow straight line is the original optimal dividing line, the red anchor point is the end point of the cut repair, and the green curve is the closed edge after completion, realizing the complete morphological repair of the radial structure.

[0046] Optionally, step S4 includes the following steps: Step S41: Determine the two intersection points of the optimal segmentation line and the edge contour of the biological adhesion mass to be treated as anchor points for incision repair; Step S42: Extract the tangent slope of the two anchor points in the contour neighborhood of the biological adhesion mass to be treated, as the boundary derivative constraint condition for incision repair; Step S43: Generate a smooth transition curve between the two anchor points by combining the boundary derivative constraints. Step S44: Replace the straight-line optimal segmentation line with a smooth transition curve to form a closed edge of a single organism, thereby obtaining a single organism image.

[0047] In one embodiment, determining the two intersection points of the optimal segmentation line and the edge contour of the biological adhesion mass to be processed as anchor points for incision repair specifically involves: firstly, obtaining the set of coordinates of all pixels traversed by the optimal segmentation line, which is generated by the Bresenham line algorithm; simultaneously, obtaining the set of pixel coordinates of the edge contour of the biological adhesion mass to be processed; performing an intersection operation on the two coordinate sets to filter out pixels with overlapping spatial positions; if the discreteness of the digital image results in more than two overlapping points, then calculating the distance between the overlapping points, selecting the two endpoints with the greatest distance as anchor points for incision repair, and marking them as the first anchor point and the second anchor point, respectively.

[0048] Specifically, the procedure for extracting the tangent slope of the two anchor points within the contour neighborhood of the bioadhesive cluster to be processed is as follows: Taking the first anchor point as an example, along the direction of the preserved single bioadhesive contour, a preset number of neighboring pixels are selected outward, for example, five adjacent contour pixels before and after the first anchor point; a coordinate matrix is ​​constructed, and the least squares method is used to perform local straight line fitting on these five pixels and the first anchor point to obtain the equation of the fitted straight line, and the slope value of the straight line is extracted; the slope value is defined as the boundary derivative constraint condition at the first anchor point.

[0049] In one implementation of this invention, generating a smooth transition curve between two anchor points by combining boundary derivative constraints specifically includes: generating a smooth curve using a cubic Bézier curve algorithm, which requires determining four control points, wherein the first control point and the fourth control point are the aforementioned first anchor point and the second anchor point, respectively; the second control point and the third control point are calculated based on the boundary derivative constraints.

[0050] It should be noted that the calculation methods for the second and third control points are as follows: First, calculate the Euclidean distance between the first and second anchor points, denoted as the chord length; multiply the chord length by a preset smoothing coefficient (e.g., 0.33) to obtain the control arm length; determine the tangent direction based on the tangent slope at the first anchor point, and extend a distance of one control arm length from the first anchor point along the tangent direction into the segmented area; this position is the second control point; similarly, extend the same length from the second anchor point along its tangent direction to obtain the third control point.

[0051] In one embodiment, the grayscale value of the straight line pixel corresponding to the original optimal segmentation line is set to the background value 0; the coordinates of the generated smooth transition curve are written into the image matrix and assigned the foreground value 255, thereby connecting the first anchor point and the second anchor point; finally, the closed area is filled with holes to form a single biological binary mask, and the single biological image is extracted from the in-situ microscopic image using the mask.

[0052] Of particular importance, step S4, which utilizes the local tangent features of the edge contour to smoothly complete the separated cut edges, also includes: The bio-agglomerates to be treated are separated into independent sub-regions along the optimal segmentation line; The two breakpoints that determine the optimal dividing line and the outline of the separated independent sub-regions are defined as the cut repair endpoints; Along the contour of the independent sub-region, extend a preset length outward to the two incision repair endpoints respectively, and extract the contour curvature change rate sequence within the extended segment; An edge extension prediction model is established based on the contour curvature change rate sequence. The natural closing trend of biological edges is simulated along the tangent direction of the incision repair endpoint to generate a virtual predicted boundary curve. Calculate the area of ​​the compensation region enclosed by the virtual predicted boundary curve and the optimal segmentation line; if the ratio of the compensation region area to the area of ​​the independent sub-region is less than the preset deformation threshold, then replace the optimal segmentation line with the virtual predicted boundary curve to complete the closure repair of the cut edge, thereby obtaining a single biological image.

[0053] Specifically, the operation procedure for extracting the contour curvature change rate sequence within the extended segment along the contour of an independent sub-region is as follows: Starting from the initial endpoint, along the edge of the biological contour away from the incision direction, select a set of pixels of a preset length, and calculate the curvature value of each pixel in the set using the discrete curvature formula, which is: ; in, , The first derivative of the pixel coordinates. , It is the second derivative. The curvature value is calculated; the difference in curvature values ​​between adjacent pixels is used to form a sequence of contour curvature change rates.

[0054] In one implementation of this invention, establishing an edge extension prediction model based on the contour curvature change rate sequence specifically includes: constructing a cubic polynomial regression model using the least squares method; using the coordinates of the extracted extension segment pixels as training data, performing interpolation prediction within the horizontal coordinate range between the starting endpoint and the ending endpoint to generate coordinate points that conform to the original bending trend of the edge, and connecting these coordinate points to form a virtual prediction boundary curve.

[0055] Optionally, step S4 may further include the following steps: Step S45: Extract the sub-region area and convex hull fill rate of the single biological image; Step S46: Compare the area of ​​the sub-region with the preset single-unit area threshold, and compare the convex hull filling rate with the preset compactness threshold. Step S47: If the area of ​​the sub-region is greater than the preset single-unit area threshold and the convex hull filling rate is lower than the preset compactness threshold, then the single-unit biological image is determined to be a residual adhesive mass. Step S48: Treat the residual adhering clumps as new biological adhering clumps to be processed, and return to steps S2 to S4 until all the output individual biological images do not meet the judgment condition of step S47.

[0056] In one embodiment, extracting the sub-region area and convex hull fill rate of a single biological image specifically involves: first, counting the total number of foreground pixels in the binary mask of the single biological image, and using this number as the sub-region area; then, using a convex hull detection algorithm to calculate the minimum convex polygon surrounding the foreground region, counting the total number of all pixels (including foreground and background) contained within the minimum convex polygon, and using this as the convex hull area.

[0057] In one implementation of this invention, when executing the determination logic of steps S46 and S47, the feature data of the currently processed single biological image is read and compared with a preset threshold.

[0058] In one embodiment, the iterative processing of step S48 is specifically as follows: when a single biological image is determined to be a residual adherent mass, it is not output as the final result, but is re-marked as a biological adherent mass to be processed, and its image data is pushed to the end of the processing queue; the system scheduler extracts the image from the queue again, calls step S2 again to extract contours and locate extreme points, and then executes steps S3 and S4 to perform segmentation again; this recursive process continues until the area of ​​all segmented sub-image blocks is less than the preset single-unit area threshold, or its convex hull filling rate is higher than the preset compactness threshold, at which point the segmentation is considered to be completely completed.

[0059] Preferably, step S5: input the individual biological image into a pre-trained classifier to obtain a biological abundance dataset containing species category, number of individuals, and volume concentration.

[0060] Optionally, step S5 includes the following steps: Step S51: Extract the morphological representation vector of the single biological image; Step S52: Input the morphological representation vector into the pre-trained planktonic classifier to identify the specific species category of each individual organism and assign a species label; Step S53: Count the number of labels for the same species, combine the sampling volume from the in-situ microscopic images, calculate the volume concentration of each species, and generate a biodiversity dataset containing species category, number of individuals, and volume concentration.

[0061] In one embodiment, the extraction of the morphological representation vector of a single biological image is specifically as follows: First, the segmented single biological image is preprocessed by using bilinear interpolation to uniformly scale the image to a fixed pixel size and then normalizing it to map the pixel values ​​to between 0 and 1. Subsequently, the processed image is input into a deep convolutional neural network feature extraction model, which removes the original last classification layer and retains only the convolutional and pooling layers. The global average pooling output of the penultimate layer of the model is extracted and used as a one-dimensional high-dimensional array, i.e., the morphological representation vector.

[0062] Specifically, the process of inputting morphological representation vectors into a pre-trained plankton classifier and identifying categories is as follows: The pre-trained plankton classifier is constructed using a fully connected neural network layer connected to a Softmax activation function; the extracted morphological representation vectors are input into the fully connected layer, and the confidence scores corresponding to each species category are output through matrix multiplication and bias addition operations; the confidence scores are converted into a probability distribution using the Softmax function, and the category with the highest probability value is selected as the identification result; if the maximum probability value exceeds a preset confidence threshold, the name corresponding to the category is assigned to the individual organism as a species label, otherwise it is marked as an unknown category.

[0063] It should be noted that the construction and training process of the pre-trained plankton classifier is as follows: A labeled image dataset containing no fewer than 50 common marine plankton species is constructed and divided into a training set and a validation set in an 8:2 ratio; a deep learning network architecture is built, and the loss function is set as the cross-entropy loss function, which measures the difference between the predicted probability distribution and the true label distribution; the stochastic gradient descent optimization algorithm is used to iteratively update the network parameters on the training set until the loss function value converges on the validation set; finally, the trained network parameter model is saved as the pre-trained plankton classifier.

[0064] In one implementation of this invention, calculating the volume concentration of each species and generating a biodiversity dataset specifically includes: traversing the species labels of all individual biological images processed in the current batch, and using a hash table or counter to count the total number of occurrences of each species label, which is recorded as the number of individuals. The sampling volume during in-situ microscopic image capture is obtained by: reading the field-of-view area parameter and depth-of-field parameter of the imaging device, multiplying them to obtain the sampling water volume of a single frame image, and then multiplying this by the total number of frames of the currently processed in-situ microscopic images to obtain the total sampling volume.

[0065] In another embodiment of the present invention, shape descriptors are extracted from images of individual organisms; wherein, the shape descriptors include Fourier descriptors and moment invariant features; the shape descriptors are input into a pre-trained plankton classifier to identify the specific species category of each individual organism and assign a species label. In one embodiment, the specific operation process for extracting Fourier descriptors is as follows: First, extract the closed edge contour of the single biological image, and obtain the descriptors arranged clockwise on the contour. The coordinates of each pixel; Constructed in complex form ,in Let the imaginary unit be the complex number; perform a discrete Fourier transform on this complex number sequence and calculate the coefficients after the transform. The first 15 low-frequency coefficients in the transformation result are selected as shape features, and these coefficients are normalized by dividing them by the magnitude of the DC component (i.e., the first coefficient), thereby eliminating the effects of image translation, rotation and scale changes, and finally obtaining a Fourier descriptor vector with scale invariance.

[0066] In one implementation of this invention, extracting moment invariant features specifically refers to calculating the Hu moment features of the image, the steps of which include: converting the single biological image into a binary image, and calculating the Hu moment features of the image. The first central moment is calculated using the following formula: ; in, and Represents the x and y coordinates of a pixel. and Represents the centroid coordinates of the image. The gray value representing a pixel (0 or 1 in a binary image). and The values ​​are non-negative integers; the normalized central moments are calculated based on the central moments, and then seven Hu moment values ​​with translation, rotation and scaling invariance are calculated through nonlinear combination to form a 7-dimensional moment invariant eigenvector.

[0067] Specifically, inputting the shape descriptor into the pre-trained planktonic classifier for identification includes: concatenating the extracted 15-dimensional Fourier descriptor vector with the 7-dimensional moment invariant feature vector to form a 22-dimensional comprehensive shape feature vector; inputting this comprehensive shape feature vector into a pre-trained support vector machine (SVM) classifier; the SVM classifier uses a pre-constructed hyperplane to divide the high-dimensional feature space into different category regions, determines the region into which the current input vector falls, and outputs the corresponding species category index; and searching the species name mapping table according to the index, assigning the corresponding organism name to the individual organism as a species label.

[0068] Please see Figure 4This is a schematic diagram of the marine planktonic image segmentation process of the present invention. The in-situ microscopic image is the input data to be processed, containing multiple adhered biological individuals. The binarization processing module is connected to the in-situ microscopic image input end, performs threshold segmentation on the grayscale image, separates the biological targets from the background, and outputs a black and white binary image. The connected component analysis unit receives the binary image, identifies independent regions through pixel connectivity detection, and marks the adhered biological clumps to be processed.

[0069] By tracing the boundaries of the biological aggregates to be treated, and moving along the contour through a sliding sampling window, the set of extreme points of the contour depression is located; the topology graph construction unit simultaneously extracts the skeleton, and establishes a weighted undirected graph-like topology graph with pixel length as the edge weight.

[0070] The integrity score of the biological structure is obtained, the candidate paths are mapped to the topological structure graph, the intersection of the path and the skeleton is detected, and the score is calculated by combining the aspect ratio of the indentation, edge weight and interference coefficient. The path with the highest score is selected as the optimal dividing line.

[0071] The intersection of the optimal segmentation line and the edge contour is determined as the anchor point for cut repair. A smooth transition curve is generated to replace the straight segmentation line, forming a closed single-organism edge and outputting a single-organism image.

[0072] The pre-trained classifier includes a morphological feature extraction unit and a species recognition unit, which extract morphological representation vectors and identify species categories to assign labels. The biodiversity dataset generator counts the number of individuals of the same species, calculates the volume concentration based on the sampled volume, and outputs a biodiversity dataset containing species category, number of individuals, and volume concentration. The residual adhesion detector extracts the area of ​​sub-regions and the convex hull filling rate to determine whether residual adhesion exists. If it exists, it returns to the edge contour extraction module for recursive processing until all individual biological images meet the separation conditions.

[0073] The present invention also provides a marine plankton image segmentation system that performs the marine plankton image segmentation method described above. The marine plankton image segmentation system includes: The agglomerate screening module is used to acquire in-situ microscopic images of marine plankton, perform binarization processing and connected component analysis on the in-situ microscopic images, screen out target areas with an area exceeding a preset individual area threshold, and mark the target areas as agglomerates of organisms to be processed. The structural integrity analysis module is used to simultaneously extract the edge contours of the biological aggregates to be processed, locate the contour depression extreme points, and construct a topological structure map. Based on the set of contour depression extreme points, multiple candidate segmentation paths are constructed, the candidate segmentation paths are mapped to the topological structure map, and the biological structural integrity score is calculated by combining the topological fracture cost of the candidate segmentation paths. The optimal segmentation decision module is used to select the candidate segmentation path with the highest biological structure integrity score as the optimal segmentation line. The single-organism separation and repair module is used to separate the biological adhesion clumps to be processed into independent sub-regions along the optimal segmentation line, and to perform curvature smoothing and completion on the separated cut edges to obtain single-organism images; The biological classification and statistics module is used to input individual biological images into a pre-trained classifier to obtain a biological abundance dataset that includes species category, number of individuals, and volume concentration.

[0074] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0075] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for segmenting images of marine plankton, characterized in that, Includes the following steps: Step S1: Obtain in-situ microscopic images of marine plankton, perform binarization processing and connected component analysis on the in-situ microscopic images, filter out target areas with an area exceeding the preset individual area threshold, and mark the target areas as biological aggregates to be processed. Step S2: Simultaneously extract the edge contour of the biological adhesion mass to be processed to locate the contour concavity extreme points and construct a topological structure map at the same time; construct multiple candidate segmentation paths based on the set of contour concavity extreme points, map the candidate segmentation paths to the topological structure map, and calculate the biological structure integrity score by combining the topological fracture cost of the candidate segmentation paths. Step S3: Select the candidate segmentation path with the highest biological structure integrity score as the optimal segmentation line; Step S4: Separate the biological adhesion clumps to be processed into independent sub-regions along the optimal segmentation line, and perform curvature smoothing to complete the separated cut edges to obtain individual biological images; Step S5: Input the individual biological image into the pre-trained classifier to obtain a biological abundance dataset containing species category, number of individuals, and volume concentration.

2. The method for image segmentation of marine plankton according to claim 1, characterized in that, Step S2 involves simultaneously extracting the edge contour of the bio-adhesive aggregates to be processed in order to locate the extreme points of the contour depressions, including: The covariance matrix of all pixel coordinates within the window is calculated by moving the preset sliding sampling window along the edge contour of the biological adhesion mass to be processed. The discrete curvature value of the center pixel of the window is determined based on the ratio of eigenvalues ​​of the covariance matrix, and the curvature distribution curve of the whole contour is generated. The extreme value search is performed on the curvature distribution curve of the whole contour to identify the trough position where the curvature value is lower than the adjacent pixel point, and it is defined as the local curvature minimum point. Set a negative curvature threshold, remove fluctuations with curvature values ​​greater than the negative curvature threshold from the local curvature minima, and determine the remaining local curvature minima as contour concavity extreme points.

3. The method for segmenting marine planktonic images according to claim 1, characterized in that, Step S2, which simultaneously constructs the topology diagram, includes: The biological adhesion clumps to be treated are peeled off layer by layer until a central skeleton line with a single pixel width is retained. Calculate the eight-neighbor connectivity of each pixel on the central skeleton line, mark points with a connectivity greater than two as skeleton branch points, and mark points with a connectivity of one as skeleton endpoints; Tracing the connection paths between skeleton branch points, a weighted undirected graph is constructed with skeleton branch points and skeleton endpoints as nodes and connection paths as edges. The pixel length of the connection path is used as the edge weight of the weighted undirected graph to form a topological structure graph.

4. The method for segmenting marine planktonic images according to claim 1, characterized in that, Step S2 involves constructing multiple candidate segmentation paths based on the set of extreme points of the contour concavity, including: Calculate the Euclidean distance between all contour concave extreme points, and pair non-adjacent contour concave extreme points with a distance less than the preset adhesion width to form candidate segmentation point pairs; Connect candidate split point pairs to form a straight line segment, or extend along the normal direction of the two point pairs to intersect and form a broken line segment. Define the straight line segment or broken line segment as a candidate split path.

5. The method for segmenting marine plankton images according to claim 4, characterized in that, Before defining straight line segments or polyline segments as candidate segmentation paths, the validity of the pairings is also verified: For the two contour concavity extreme points in the candidate segmentation point pair, search for the pixel point with the closest Euclidean distance on the central skeleton line of the topology map, and use it as the skeleton projection point. Search for the shortest skeleton path between two skeleton projection points in the topology graph and calculate the topological distance of the shortest skeleton path. Calculate the Euclidean distance between two extreme points of the contour concavity. If the ratio of the Euclidean distance to the topological distance is less than the preset path curvature threshold, the candidate segmentation point pair is determined to be valid and candidate segmentation paths are allowed to be generated; otherwise, it is determined to be an invalid pair.

6. The method for segmenting marine plankton images according to claim 1, characterized in that, Step S2 maps candidate segmentation paths to a topological graph and calculates the biological structure integrity score by combining the topological fracture cost of the candidate segmentation paths, including: The depth-to-width ratio feature of the concave is calculated based on the extreme points of the concave contours at both ends of the candidate segmentation path. Detect whether the candidate segmentation path intersects with the connection path in the topology graph, and mark the skeleton edge at each intersection position as a topological interference edge; Assign topological interference coefficients based on the node types at both ends of the topological interference edge; The original edge weights of the topological interference edges are obtained from the topological structure graph, and the biological structure integrity score of the candidate segmentation path is calculated based on the indentation aspect ratio, the original edge weights and the topological interference coefficient.

7. The method for segmenting marine plankton images according to claim 1, characterized in that, Step S4 includes the following steps: Step S41: Determine the two intersection points of the optimal segmentation line and the edge contour of the biological adhesion mass to be treated as anchor points for incision repair; Step S42: Extract the tangent slope of the two anchor points in the contour neighborhood of the biological adhesion mass to be treated, as the boundary derivative constraint condition for incision repair; Step S43: Generate a smooth transition curve between the two anchor points by combining the boundary derivative constraints. Step S44: Replace the straight-line optimal segmentation line with a smooth transition curve to form a closed edge of a single organism, thereby obtaining a single organism image.

8. The method for segmenting marine plankton images according to claim 7, characterized in that, Step S4 also includes the following steps: Step S45: Extract the sub-region area and convex hull fill rate of the single biological image; Step S46: Compare the area of ​​the sub-region with the preset single-unit area threshold, and compare the convex hull filling rate with the preset compactness threshold. Step S47: If the area of ​​the sub-region is greater than the preset single-unit area threshold and the convex hull filling rate is lower than the preset compactness threshold, then the single-unit biological image is determined to be a residual adhesive mass. Step S48: Treat the residual adhering clumps as new biological adhering clumps to be processed, and return to steps S2 to S4 until all the output individual biological images do not meet the judgment condition of step S47.

9. The method for segmenting marine plankton images according to claim 1, characterized in that, Step S5 includes the following steps: Step S51: Extract the morphological representation vector of the single biological image; Step S52: Input the morphological representation vector into the pre-trained planktonic classifier to identify the specific species category of each individual organism and assign a species label; Step S53: Count the number of labels for the same species, combine the sampling volume from the in-situ microscopic images, calculate the volume concentration of each species, and generate a biodiversity dataset containing species category, number of individuals, and volume concentration.

10. A system for segmenting images of marine plankton, characterized in that, For performing the marine plankton image segmentation method as described in claim 1, the marine plankton image segmentation system comprises: The agglomerate screening module is used to acquire in-situ microscopic images of marine plankton, perform binarization processing and connected component analysis on the in-situ microscopic images, screen out target areas with an area exceeding a preset individual area threshold, and mark the target areas as agglomerates of organisms to be processed. The structural integrity analysis module is used to simultaneously extract the edge contours of the biological aggregates to be processed, locate the contour depression extreme points, and construct a topological structure map. Based on the set of contour depression extreme points, multiple candidate segmentation paths are constructed, the candidate segmentation paths are mapped to the topological structure map, and the biological structural integrity score is calculated by combining the topological fracture cost of the candidate segmentation paths. The optimal segmentation decision module is used to select the candidate segmentation path with the highest biological structure integrity score as the optimal segmentation line. The single-organism separation and repair module is used to separate the biological adhesion clumps to be processed into independent sub-regions along the optimal segmentation line, and to perform curvature smoothing and completion on the separated cut edges to obtain single-organism images; The biological classification and statistics module is used to input individual biological images into a pre-trained classifier to obtain a biological abundance dataset that includes species category, number of individuals, and volume concentration.

Images