Image feature classification method for hydrophilic fiber dressing density

Abstract

The application discloses a hydrophilic fiber dressing density image feature classification method, and belongs to the technical field of image processing and material detection. The method comprises the following steps: obtaining a multi-scale microscopic image sequence obtained by layering scanning of a hydrophilic fiber dressing sample by changing an imaging focal length; performing a fiber trajectory tracking operation on the multi-scale microscopic image sequence, extracting spatial distribution coordinates of fiber cross nodes and a continuous change path of fiber orientation angles; constructing a fiber network topology diagram across layers according to the spatial distribution coordinates, marking key nodes crossed by the continuous change path of the fiber orientation angles on the diagram, and generating a density feature backbone network; dividing the backbone network into a plurality of local analysis windows, calculating the key node aggregation degree and the fiber orientation angle dispersion degree in each window, and obtaining a local classification feature vector; and performing weighted voting on the local classification feature vectors of all local analysis windows according to a spatial adjacent relationship, and outputting an overall density category label of the hydrophilic fiber dressing sample.

Description

Technical Field

[0001] This invention relates to the fields of image processing and material testing technology, specifically to an image feature classification method for the density of hydrophilic fiber dressings. Background Technology

[0002] The density of hydrophilic fiber dressings is a crucial indicator of their absorbency, structural stability, and clinical performance. Accurate density classification is essential for quality control and product selection. Existing dressing density classification methods largely rely on single-focal-length or fixed-focal-plane microscopic images, utilizing grayscale texture statistics, porosity analysis, or direct density estimation based on image pixel intensity. These approaches treat fiber distribution as a texture feature in a two-dimensional plane, establishing a mapping relationship between features and density levels by calculating the local grayscale co-occurrence matrix, fractal dimension, or fiber coverage. Some methods also employ traditional image segmentation to extract fiber regions and measure their area proportions to represent the overall dressing density.

[0003] Existing technical solutions have significant drawbacks. On the one hand, single-layer imaging loses information about the intersections and connections of fibers in the depth direction, failing to express the permeability of the fiber network between different layers, resulting in a severely distorted description of the fiber's three-dimensional structure. When fibers bend, intersect, and entangle at different levels, single-layer texture alone cannot distinguish between dense areas and simple fiber accumulation areas, easily leading to misjudgments of density levels. On the other hand, static image features struggle to capture the continuous changes in fiber path direction, especially lacking the ability to identify key density-bearing nodes such as abrupt changes in fiber angle or local twists. This makes density estimation results more susceptible to interference from fiber thickness variations, illumination fluctuations, and local background noise, resulting in insufficient overall classification robustness.

[0004] The problem this method aims to solve is how to extract the core structure that can characterize the tightness of fiber interweaving from multi-scale layered microscopic images, construct a backbone network that reflects the density distribution in three-dimensional space, and use the local statistical differences on the backbone network to make a highly reliable density category determination. Summary of the Invention

[0005] To address the problem that existing hydrophilic fiber dressing density classification technologies suffer from inaccurate density level determination due to neglecting fiber three-dimensional spatial orientation and intersection information, the present invention aims to provide an image feature classification method for hydrophilic fiber dressing density. This method extracts the continuous change paths of fiber intersections and orientation angles from multi-scale microscopic images, constructs a backbone network that highlights the core density region, and, based on this, achieves accurate classification of the overall density category of the dressing by comprehensively utilizing the statistical characteristics and spatial adjacency relationships of local analysis windows.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides an image feature classification method for the density of hydrophilic fiber dressings. This method includes the following steps: acquiring a multi-scale microscopic image sequence, which is a set of original images obtained by layer-by-layer scanning of a hydrophilic fiber dressing sample with varying imaging focal lengths. This allows for the acquisition of fiber morphology information at different depth levels without damaging the internal structure of the sample. Performing fiber trajectory tracking on the multi-scale microscopic image sequence allows for the extraction of the spatial distribution coordinates of fiber intersection nodes and the continuous change path of fiber orientation angles in each layer of the microscopic image, thereby transforming the microscopic interweaving state of the fiber network into quantifiable spatial and directional features. Based on the spatial distribution coordinates of the fiber intersection nodes, a cross-layer fiber network topology is constructed, and key nodes traversed by the continuous change path of fiber orientation angles are marked on this fiber network topology, generating a density feature backbone network. This density feature backbone network eliminates secondary fiber connections while retaining the structural skeleton with high orientation abrupt changes, thus more effectively reflecting the essential characteristics of dressing density differences. The density feature backbone network is divided into multiple local analysis windows, and the clustering degree of key nodes and the dispersion of fiber orientation angles within each local analysis window are calculated to obtain local classification feature vectors. This transforms the global density evaluation into a feature expression with spatial locality, improving its adaptability to non-uniform density distributions. The local classification feature vectors of all local analysis windows are weighted according to spatial adjacency to output the overall density category label of the hydrophilic fiber dressing sample. The spatial dependency between adjacent windows is used to smooth the classification results, improving the accuracy and robustness of the overall density classification.

[0007] As a preferred embodiment of the present invention, the fiber trajectory tracking operation specifically includes: applying gradient direction histogram matching to the current layer microscopic image to identify the fiber pixel set, and selecting pixels whose curvature changes exceed a set threshold as candidate intersection points; determining the points in the eight-neighborhood of the candidate intersection points that have fiber extensions in at least three directions as fiber intersection nodes, and recording their spatial distribution coordinates in the image coordinate system; refining the fiber segment between two adjacent fiber intersection nodes using skeleton lines, calculating the tangent direction pixel by pixel along the skeleton lines, and using the angle sequence of all obtained tangent directions as the continuous change path of the fiber orientation angle. This process converts discrete fiber pixel information into a continuous node and path representation, providing accurate basic data for subsequent topology analysis and backbone network extraction.

[0008] In constructing the density feature backbone network, a preferred implementation method is as follows: The spatial distribution coordinates of the fiber cross nodes extracted from each layer are mapped to a unified three-dimensional coordinate system. Between adjacent microscopic images of different layers, node connections between layers are established by comparing the angle between the endpoint and starting points of the continuous change path of fiber orientation angles. A cross-layer fiber network topology graph is generated, with fiber cross nodes as vertices and node connections as edges. Each edge in this topology graph is traversed, and it is determined whether there is a sudden change point where the angle jump exceeds a threshold on the continuous change path of the fiber orientation angle between the two fiber cross nodes corresponding to the edge. If such a point exists, the fiber cross node corresponding to the sudden change point is marked as a key node. The density feature backbone network is composed of all key nodes and their connecting edges. Key nodes in this backbone network are assigned higher topological connectivity weights due to their corresponding positions of drastic fiber orientation changes, thereby strengthening the expression of density-sensitive locations in the feature space. The specific process for establishing interlayer connections is as follows: Obtain the endpoint direction angle of the continuous path of fiber orientation angle change corresponding to a fiber cross node in the previous layer, and the starting direction angle of the continuous path of change corresponding to a fiber cross node in the current layer. Calculate the absolute difference between the two. When the absolute difference is less than the angle connectivity threshold, create an interlayer connection between the two fiber cross nodes. Repeat this comparison process, traversing all fiber cross node pairs in all adjacent layers to complete the establishment of interlayer node connections. When marking key nodes on each edge of the topology graph, select an edge from the cross-layer fiber network topology graph, obtain the two fiber cross nodes it connects, and extract all angle sampling values ​​between the two points. Calculate the jump amplitude between two adjacent angle sampling values. If there is a sudden change point where the jump amplitude exceeds the jump threshold, mark both fiber cross nodes as key nodes. After traversing all edges, collect all fiber cross nodes marked as key nodes and delete unmarked nodes and their connecting edges to obtain the density feature backbone network. This process enables the compression and abstraction of multi-layered fiber interwoven structures, allowing the density-featured backbone network to compactly and accurately characterize the tightness and interweaving morphology of the fiber network inside the dressing.

[0009] As a further improvement of the present invention, the step of dividing the density feature backbone network into multiple local analysis windows and calculating local classification feature vectors includes: setting a three-dimensional sliding window of fixed size on the density feature backbone network, moving the three-dimensional sliding window according to a preset step size, and taking the area covered by the window after each sliding as a local analysis window; for each local analysis window, counting the total number of key nodes contained in the window, and dividing the total number by the window volume to obtain the degree of aggregation of the key nodes; extracting all directional angles on the continuous change path of the fiber orientation angles corresponding to all connecting edges in the window, calculating the circular variance of all directional angles, and taking the circular variance as the dispersion of the fiber orientation angles; combining the degree of aggregation of the key nodes and the dispersion of the fiber orientation angles into a two-dimensional vector as the local classification feature vector. By utilizing the two complementary local features of aggregation degree and orientation dispersion, the dressing density can be finely characterized from two dimensions: node compactness and fiber orientation orderliness, effectively distinguishing between high-density compact areas and low-density loose areas.

[0010] For the output of the overall density category label, the local classification feature vector of each local analysis window is input into a pre-trained window-level classifier. This classifier outputs the probability distribution of the window belonging to different density categories. During training, multiple hydrophilic fiber dressing samples of different density categories are collected. Local classification feature vectors for all local analysis windows are extracted using the aforementioned method, and the overall density category of the sample to which each local classification feature vector belongs is labeled, thus constructing a training dataset. A support vector machine (SVM) model is used as the basic framework for training the window-level classifier, enabling the trained SVM model to output the probability value of each local analysis window belonging to different density categories. In the classification stage, the set of neighboring windows is determined based on the spatial position of the center point of each local analysis window in the density feature backbone network. Neighboring windows satisfy the condition that the distance between them and the local analysis window in three-dimensional space is less than one window size. For each density category, the corresponding probability value in the window category probability of that local analysis window is weighted and averaged with the corresponding probability values ​​of all adjacent windows. The weighting coefficient is inversely proportional to the distance between adjacent windows, resulting in a weighted voting score for that local analysis window. The weighted voting scores of all local analysis windows are then summed, and the density category with the highest summation score is selected as the overall density category label. This weighted voting process integrates spatial context information, suppresses isolated classification noise, and ensures the stability of the overall discrimination result. Preferably, the weighting coefficients of different local analysis windows can also be adaptively adjusted based on the distance from the center point of the window to the geometric center of the hydrophilic fiber dressing sample, so that the fiber structure features of the sample's central region contribute more to the final classification result, further aligning with the characteristic that the central region is more representative in actual dressing use.

[0011] When acquiring multi-scale microscopic image sequences, hydrophilic fiber dressing samples are fixed on the stage of a microscopic imaging platform. A vertical scanning range from the upper to the lower surface of the dressing sample is set, and the vertical scanning range is divided into multiple imaging layers according to a fixed interlayer spacing. At each imaging layer, the autofocus system is controlled to adjust the imaging focal length to the depth of that layer, triggering the camera to acquire the corresponding raw microscopic image, and simultaneously recording the depth coordinate value of the raw microscopic image. All acquired raw microscopic images are arranged into a sequence according to the ascending order of depth coordinate values, and illumination non-uniformity correction and noise filtering preprocessing are performed to obtain a high-quality multi-scale microscopic image sequence, laying a reliable data foundation for subsequent fiber tracking and feature extraction.

[0012] Through the above scheme, this invention extracts the cross-layer fiber network topology and its backbone network from multilayer microscopic images of hydrophilic fiber dressings. By utilizing the spatial aggregation and orientation dispersion features based on local windows, combined with the weighted voting mechanism of adjacent windows, it achieves high-precision and robust automatic classification of dressing density, solving the problem that traditional visual inspection or simple image analysis methods cannot accurately reflect the density differences of three-dimensional interwoven structures.

[0013] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0014] By acquiring multi-scale microscopic image sequences and performing fiber trajectory tracing on each layer of microscopic images, the spatial distribution coordinates of fiber intersection nodes and the continuous change path of fiber orientation angles are extracted. Based on this, a cross-layer fiber network topology map is constructed. During the generation of the topology map, abrupt changes in angle exceeding a threshold are detected along each edge on the continuous change path of fiber orientation angles. If such abrupt changes occur, the fiber intersection node connected to that edge is marked as a key node. All key nodes and their connecting edges form a density feature backbone network. This processing method allows the densified regions formed by sharp turns and intertwining of fiber orientations in the fiber network to be extracted separately, eliminating redundant information in straight and continuous fiber regions. The backbone network directly corresponds to the high-connectivity structure inside the dressing that truly determines the density level, effectively solving the problem of insensitivity of traditional two-dimensional texture analysis to fiber spatial configuration and significantly improving the feature discrimination of density classification.

[0015] The density feature backbone network is divided into multiple local analysis windows. Within each window, the circular variance of the clustering degree of key nodes and fiber orientation angles is statistically analyzed, forming a local classification feature vector composed of clustering degree and dispersion. The window category probabilities are then weighted and voted on based on the adjacency relationship of the windows in three-dimensional space, ultimately outputting the overall density category label. The clustering degree of key nodes reflects the local structural density of fibers, while the dispersion of fiber orientation angles quantifies the degree of disorder in fiber arrangement. These two indicators characterize density properties from different perspectives and are highly complementary. The weighted voting mechanism ensures that the classification results of local windows are constrained by the information of neighboring windows; windows that are closer together have a greater weight. This smooths out misjudgments caused by local anomalies in fiber distribution within a single window, ensuring a more stable and accurate determination of the overall density category. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0017] Figure 1 This is a flowchart of an image feature classification method for the density of hydrophilic fiber dressings;

[0018] Figure 2 This is a flowchart of fiber cross node extraction and orientation angle measurement.

[0019] Figure 3 This is a flowchart of a method for microscopic imaging and multi-scale image sequence construction of hydrophilic fiber dressing samples;

[0020] Figure 4 This is a schematic diagram of the continuous change path of the fiber orientation angle;

[0021] Figure 5 This is a schematic diagram of key nodes and interlayer connection edges in a cross-layer fiber network topology diagram;

[0022] Figure 6 This is a density category distribution diagram of the aggregation degree of key nodes and the dispersion of fiber orientation angle in hydrophilic fiber dressings. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] See Figure 1 This invention provides an image feature classification method for the density of hydrophilic fiber dressings, comprising: acquiring a multi-scale microscopic image sequence, wherein the multi-scale microscopic image sequence is a set of original images obtained by performing layered scanning of a hydrophilic fiber dressing sample by changing the imaging focal length; performing fiber trajectory tracking operation on the multi-scale microscopic image sequence to extract the spatial distribution coordinates of fiber intersection nodes and the continuous change path of fiber orientation angle in each layer of microscopic image; constructing a cross-layer fiber network topology map based on the spatial distribution coordinates of the fiber intersection nodes, and marking the key nodes traversed by the continuous change path of fiber orientation angle on the fiber network topology map to generate a density feature backbone network; dividing the density feature backbone network into multiple local analysis windows, and calculating the clustering degree of key nodes and the dispersion of fiber orientation angle in each local analysis window to obtain a local classification feature vector; and weighting the local classification feature vectors of all local analysis windows according to spatial adjacency to output the overall density category label of the hydrophilic fiber dressing sample.

[0025] Example 1:

[0026] In specific implementation, please refer to Figure 2 Gradient orientation histogram (HOR) matching is applied to the current layer microscopic image in a multi-scale microscopic image sequence to identify the fiber pixel set. The gradient magnitude and direction of each pixel in the current layer microscopic image are calculated. The gradient direction range from 0 to 180 degrees is divided into several direction intervals, each covering the same angle range. For each pixel, its gradient magnitude is accumulated into the direction interval corresponding to its gradient direction to construct the gradient orientation histogram of the pixel's neighborhood region. A set of template images containing typical fiber textures is pre-prepared, and a standard gradient orientation histogram template is calculated for each template image. The correlation between the gradient orientation histogram of the neighborhood region of each pixel in the current layer microscopic image and all standard gradient orientation histogram templates is compared. When the correlation coefficient exceeds a preset matching threshold, the pixel is marked as a fiber pixel, and all marked fiber pixels form the fiber pixel set.

[0027] In practice, pixels whose curvature changes exceed a set threshold are selected as candidate intersection points from the fiber pixel set. Connectivity analysis is performed on the fiber pixel set to extract pixel chains for each continuous fiber segment. For each pixel in each pixel chain, the curvature value at that pixel is calculated. The curvature value is calculated by fitting a local arc using the coordinates of the pixel and its preceding and following adjacent pixels, and taking the reciprocal of the arc radius as the curvature value. The pixel chain is scanned point by point, and the absolute difference in curvature values ​​between adjacent pixels is calculated. When the absolute difference in curvature at a pixel exceeds a set threshold, that pixel is identified as a candidate intersection point. The set threshold is predetermined based on the average bending stiffness characteristics of the fiber material.

[0028] In practice, candidate intersections with at least three fiber extension directions within their eight-neighborhood are identified as fiber intersection nodes. Each candidate intersection is centered on an eight-neighborhood of eight adjacent pixels. When any of these eight adjacent pixels belongs to a set of fiber pixels, these adjacent fiber pixels are grouped according to their connectivity, with adjacent fiber pixels in the same connectivity direction grouped into one extension direction. The number of extension directions around the candidate intersection is counted; if the number of extension directions is greater than or equal to three, the candidate intersection is determined to be a fiber intersection node. The two-dimensional pixel position coordinates of the fiber intersection node in the image coordinate system of the current layer micrograph are recorded as the spatial distribution coordinates of the fiber intersection node, where the spatial distribution coordinates include row and column coordinate values.

[0029] In practice, all fiber pixels covered by the fiber segment between two adjacent fiber intersection nodes are extracted. A morphological thinning algorithm is applied iteratively to remove edge pixels from the fiber pixels until the remaining pixels form a skeleton line with a width of one pixel. The morphological thinning algorithm uses a successive boundary removal method. In each iteration, all fiber pixels covered by the fiber segment are traversed, and edge pixels that meet preset deletion conditions are marked. The preset deletion conditions include that the pixel is not an endpoint, that deleting the pixel will not disrupt the connectivity of the fiber pixels, and that deleting the pixel will not change the topology of the fiber pixel set. After the iteration process is completed, all marked edge pixels are removed from the fiber pixel set. The above iteration process is repeated until no pixels meet the deletion conditions, and the remaining fiber pixels form a skeleton line with a width of one pixel.

[0030] In practice, starting from one of two adjacent fiber intersection nodes, each skeleton pixel is visited sequentially along the extension direction of the skeleton line. At each skeleton pixel, the position coordinates of the preceding and following adjacent skeleton pixels in the access sequence along the skeleton line are obtained. The direction of the line connecting the preceding and following adjacent skeleton pixels is taken as the tangent direction at that skeleton pixel, and the angle between the tangent direction and the horizontal axis of the image is calculated. The horizontal axis of the image is defined as a coordinate axis parallel to the pixel row direction of the current layer of the microscopic image. The angle is calculated by taking the directed angle value between the tangent direction vector and the positive direction vector of the horizontal axis of the image, and the angle value is limited to the range of 0 degrees to 180 degrees.

[0031] Optionally, the included angles corresponding to all skeleton pixels on the skeleton line are arranged into an angle array in the order they are visited along the skeleton line. The position index of each element in the angle array corresponds to the arrangement order of the skeleton pixels on the skeleton line, and the value of each element corresponds to the angle between the tangent direction at that skeleton pixel and the horizontal axis of the image. The angle array is used as a continuous variation path of fiber orientation angles, which fully represents the point-by-point change of the orientation angle of the fiber segment from one node to the other between two adjacent fiber intersection nodes along the spatial extension direction.

[0032] See Figure 4 In the figure, the horizontal axis represents the arrangement order of skeleton pixels on the skeleton line, with values ​​ranging from 0 to 600. The vertical axis represents the fiber orientation angle at the corresponding skeleton pixel, in degrees, ranging from 0 to 180 degrees. This curve depicts the continuous change path of the fiber orientation angle at each skeleton pixel in the access sequence along the skeleton line starting from one of the two adjacent fiber intersection nodes.

[0033] The curve generally shows a relatively stable and continuous trend. However, between the arrangement of the skeleton pixels (approximately 200 to 250), the fiber orientation angle rapidly jumps from about 90 degrees to nearly 180 degrees, indicating a significant bend or twist in this fiber segment. Subsequently, the fiber orientation angle fluctuates around 180 degrees, showing the fibers tending towards a more consistent direction. However, between the arrangement of pixels (approximately 400 to 600), the fiber orientation angle gradually decreases, dropping from 180 degrees to approximately 120 degrees, demonstrating a slow change in fiber orientation.

[0034] This curve reflects the subtle and continuous change in the orientation angle of the fibers at the skeleton line in Example 1 along the spatial extension direction, providing crucial directional data support for subsequently constructing a cross-layer fiber network topology and density characteristic backbone network through the continuous change path of fiber orientation angle. The curve exhibits good continuity without drastic discontinuities or jumps, consistent with the description of fiber trajectory tracking and angle calculation in Example 1, demonstrating the spatial continuity and orientation change characteristics of the fiber structure.

[0035] Example 2:

[0036] In practice, the spatial distribution coordinates of the fiber cross nodes extracted from each layer of microscopic images are mapped to a unified three-dimensional spatial coordinate system. The three-dimensional spatial coordinate system uses the stage plane of the microscopic imaging platform as the reference plane, a preset marker point on the stage plane as the origin, a direction parallel to the stage plane and parallel to the pixel row direction of the first original microscopic image in the multi-scale microscopic image sequence as the positive X-axis, a direction parallel to the stage plane and parallel to the pixel column direction as the positive Y-axis, and a direction perpendicular to the stage plane and pointing towards the upper surface of the hydrophilic fiber dressing sample as the positive Z-axis. For each fiber cross node in each layer of microscopic images, the row and column coordinate values ​​of the fiber cross node in the image coordinate system of the current layer of microscopic images are obtained. Based on the pixel size calibration coefficient of the microscopic imaging system, the row and column coordinate values ​​are converted into X and Y coordinate values ​​in the three-dimensional spatial coordinate system. Obtain the depth coordinates corresponding to the current layer of the microscopic image. The depth coordinates are the vertical distance from the imaging focal plane to the stage plane recorded when acquiring the microscopic image of this layer. The depth coordinates are used as the Z coordinates in the three-dimensional spatial coordinate system. The complete spatial distribution coordinates of the fiber cross node in the three-dimensional spatial coordinate system consist of three components: X coordinate, Y coordinate, and Z coordinate.

[0037] In practice, the endpoint orientation angle of the continuous path of fiber orientation angle change corresponding to a fiber intersection node in the previous layer of the microscopic image, and the starting orientation angle of the continuous path of fiber orientation angle change corresponding to a fiber intersection node in the current layer of the microscopic image are obtained. The endpoint orientation angle is defined as the included angle value recorded by the last element in the angle array of the continuous path of fiber orientation angle change in the previous layer of the microscopic image. The starting orientation angle is defined as the included angle value recorded by the first element in the angle array of the continuous path of fiber orientation angle change in the current layer of the microscopic image. The absolute difference between the endpoint orientation angle and the starting orientation angle is calculated. The absolute difference is calculated according to the cyclic property of the angle, and the formula is expressed as:

[0038]

[0039] in, It represents the absolute difference between the endpoint orientation angle and the starting orientation angle, in degrees, and ranges from 0 to 90 degrees. It represents the endpoint direction angle of the continuous change path of fiber orientation angle in the previous layer of microscopic image, in degrees, with a value range of 0 to 180 degrees; It represents the starting direction angle of the continuous change path of fiber orientation angle in the current layer micrograph, in degrees, with a value range of 0 to 180 degrees; It represents the absolute value of the linear difference between the endpoint direction angle and the starting direction angle without angular cyclic correction; The function represents taking the smaller of the two values. The reason for using the cyclic property of angles in the calculation is that there is no distinction between positive and negative directions for the fiber orientation angle; 0 degrees and 180 degrees geometrically represent the same fiber orientation direction.

[0040] In practice, when the absolute difference is less than the angle connectivity threshold, an interlayer connection edge is created between the fiber cross nodes in the previous layer micrograph and the fiber cross nodes in the current layer micrograph. The angle connectivity threshold is set to 25 degrees, based on the upper limit of the statistical distribution of the maximum allowable bending angle of hydrophilic fiber materials between adjacent imaging layers. When the absolute difference in fiber orientation between adjacent layers exceeds 25 degrees, it is determined that the two fiber cross nodes do not belong to the continuous extension path of the same fiber in space. The process of obtaining the endpoint orientation angle, obtaining the starting orientation angle, calculating the absolute difference, determining whether the absolute difference is less than the angle connectivity threshold, and creating the interlayer connection edge is repeated, traversing all fiber cross node pairs in all adjacent layer micrographs to complete the establishment of the node connection relationship between layers. An adjacent layer micrograph pair refers to any two layer micrographs whose depth coordinate values ​​differ by a fixed layer spacing. For each pair of adjacent layer micrographs, each fiber cross node in the previous layer micrograph is paired with each fiber cross node in the current layer micrograph in sequence, and the interlayer connectivity judgment is performed.

[0041] In practice, a cross-layer fiber network topology is generated using all fiber crossover nodes as vertices and all inter-layer connection edges as edges. Each vertex stores the X, Y, and Z coordinates of the corresponding fiber crossover node, as well as the layer number of the fiber crossover node. Each edge stores the index of the fiber crossover nodes corresponding to its two endpoints and the length of the fiber segment represented by the edge in three-dimensional space. The length of the fiber segment in three-dimensional space is calculated using the Euclidean distance between the spatial coordinates of the two fiber crossover nodes. The cross-layer fiber network topology is stored using an adjacency list data structure. Each vertex corresponds to a list of adjacent vertices, which records the indexes of all other vertices that have inter-layer connection edges with that vertex.

[0042] In practice, each edge in the cross-layer fiber network topology is traversed to determine if there are any abrupt changes in angle exceeding a threshold on the continuous path of fiber orientation angle change between the two fiber cross nodes corresponding to the edge. An edge is selected from the cross-layer fiber network topology to obtain the first and second fiber cross nodes connected by the edge. The layers where the first and second fiber cross nodes are located are designated as adjacent layers. All angle sample values ​​on the continuous path of fiber orientation angle change from the first to the second fiber cross node are extracted; these angle sample values ​​represent all elements in the angle array corresponding to the continuous path. When the first fiber cross node is located in the upper layer of the microscopic image, the angle sample values ​​of the path segment extending from the first fiber cross node to the upper layer of the microscopic image are concatenated with the angle sample values ​​of the path segment extending to the current layer of the microscopic image, forming a complete sequence of angle sample values ​​for the cross-layer continuous path. The jump amplitude between two adjacent angle sample values ​​in the angle sample value sequence is calculated, using the same angle cyclic correction calculation method as the absolute difference. When a jump amplitude exceeds the jump threshold, a sudden change point is determined to exist on the continuous change path, and both the first and second fiber intersection nodes are marked as critical nodes. The jump threshold is set to 35 degrees, based on the maximum allowable angle abrupt change tolerance of hydrophilic fibers in the extension path of a single fiber.

[0043] In practice, after traversing all edges in the cross-layer fiber network topology, all fiber cross nodes marked as key nodes are collected. Fiber cross nodes not marked as key nodes and their connecting edges are deleted from the cross-layer fiber network topology. The deletion operation includes removing the vertex entry corresponding to the fiber cross node not marked as a key node from the adjacency list and deleting the index number associated with the deleted vertex from the adjacent vertex lists of all other vertices. The remaining fiber cross nodes marked as key nodes and the retained connecting edges between these key nodes constitute the density feature backbone network.

[0044] In practical implementation, in the density feature backbone network, fiber crossing nodes corresponding to mutation points where the angle jump exceeds the jump threshold are assigned higher topological connectivity weights. For each key node in the density feature backbone network, the topological connectivity of the key node is calculated, which is the total number of connection edges directly connected to the key node in the density feature backbone network. Whether a key node corresponds to at least one mutation point where the angle jump exceeds the jump threshold is determined by whether the key node is marked during the retrieval and marking process due to a jump amplitude exceeding the jump threshold on at least one continuous path of fiber orientation angle change. When a key node corresponds to at least one mutation point where the angle jump exceeds the jump threshold, the topological connectivity of the key node is multiplied by a connectivity enhancement coefficient to obtain the topological connectivity weight. The connectivity enhancement coefficient is set to 1.5, based on the fact that the key node corresponding to the mutation point represents an abnormally concentrated area of ​​fiber crossing structure in the fiber network space, and its importance in representing density features is higher than that of ordinary crossing nodes. Increasing the connectivity enhancement coefficient reflects higher structural significance in subsequent analysis and calculation. When the critical node does not correspond to any mutation point where the angle jump exceeds the jump threshold, the topological connectivity of the critical node is directly used as the topological connectivity weight, and the connectivity enhancement coefficient is regarded as 1.0.

[0045] See Figure 5 The figure shows a two-dimensional projection of the cross-layer fiber network topology in Example 2. The horizontal and vertical axes represent the X and Y coordinates of the fiber cross nodes in the three-dimensional spatial coordinate system, respectively, with a range of 0 to 100 micrometers. Hollow circles "○" mark ordinary fiber cross nodes, and solid triangles "▲" mark key nodes. Solid lines connect the layers to represent interlayer connections, reflecting the spatial continuity between fiber cross nodes in the microscopic images of adjacent layers.

[0046] As shown in the figure, ordinary nodes and critical nodes are evenly distributed in space, with fewer critical nodes than ordinary nodes. Critical nodes mark the locations of abrupt changes in the fiber network where the angle jump exceeds a threshold (35 degrees, see Example 2). Interlayer connections are relatively sparse, mainly connecting fiber intersection nodes in adjacent layers, reflecting the continuous extension path of the fibers in space. The interlayer connections between critical and ordinary nodes are interwoven. Critical nodes, due to the abrupt changes in their corresponding fiber orientation angles, are assigned higher topological connectivity weights, representing abnormally concentrated and significantly dense regions in the fiber structure.

[0047] The spatial distribution of nodes in the figure covers the entire 100-micrometer range of the X and Y coordinates, reflecting the spatial topology of the density feature backbone network within the local sample region. This backbone network provides the foundation for subsequent local analysis window division and local classification feature vector extraction, which is beneficial for accurately characterizing the spatial variation of density characteristics and fiber orientation angles of hydrophilic fiber dressing samples.

[0048] Example 3:

[0049] In practice, a fixed-size three-dimensional sliding window is set on the density feature backbone network. The three-dimensional sliding window is a cubic spatial region, with fixed side lengths in the X, Y, and Z axes. The side length in the X-axis direction is determined based on the typical fiber cross-spacing of the hydrophilic fiber dressing sample on a horizontal plane, and is set to 15 times the average fiber cross-spacing. The side length in the Y-axis direction is set equal to the side length in the X-axis direction. The side length in the Z-axis direction is determined based on the fixed interlayer spacing and the total number of imaging layers in the multi-scale microscopic image sequence, and is set to the sum of the depth ranges covered by three adjacent imaging layers, i.e., the fixed interlayer spacing multiplied by 2. The fixed interlayer spacing is the vertical distance between adjacent imaging layers set when acquiring the multi-scale microscopic image sequence. The center point of the three-dimensional sliding window moves within the three-dimensional coordinate system area occupied by the density feature backbone network according to a preset step size. The step size is set to half the side length in the X-axis direction and half the fixed interlayer spacing in the Z-axis direction. After each slide, the three-dimensional spatial region covered by the three-dimensional sliding window is used as a local analysis window, and the boundary range of the local analysis window in the three-dimensional spatial coordinate system is recorded.

[0050] In practical implementation, for each local analysis window, the total number of key nodes contained within the local analysis window is counted. All key nodes in the density feature backbone are traversed, and the X, Y, and Z coordinates of each key node are obtained. It is then determined whether the X, Y, and Z coordinates of the key node are within the boundary range of the local analysis window along the X-axis, Y-axis, and Z-axis. When all three coordinate components are within their respective boundary ranges, the key node is considered to belong to that local analysis window. The total number of key nodes belonging to that local analysis window is obtained by summing the total number of key nodes, denoted as . .

[0051] In practice, the clustering degree of key nodes is obtained by dividing the total number of key nodes by the volume of the local analysis window. The volume of the local analysis window is calculated as the product of the side lengths along the X, Y, and Z axes of the 3D sliding window. The clustering degree of key nodes is denoted as... The calculation formula is:

[0052]

[0053] in, Indicates the degree of clustering of key nodes, measured in units per cubic micrometer. The range of values ​​for is real numbers greater than or equal to 0; This indicates the total number of key nodes contained within the local analysis window, expressed in units of [number]. The range of values ​​for is non-negative integers; This represents the side length of the 3D sliding window along the X-axis, in micrometers. The value is set based on the average cross spacing of the fibers in the hydrophilic fiber dressing sample; This represents the side length of the 3D sliding window along the Y-axis, in micrometers. The value of and equal; This represents the side length of the 3D sliding window along the Z-axis, in micrometers. The value of is determined by the fixed interlayer spacing and the number of imaging layers.

[0054] In practice, for each local analysis window, all directional angles along the continuous change path of the fiber orientation angle corresponding to all connecting edges within the local analysis window are extracted. All connecting edges located within the local analysis window in the density feature backbone network are retrieved. Connecting edges are considered to be within the local analysis window if either of their two endpoints (key nodes) belongs to the local analysis window, or if the connecting edge spatially crosses the boundary of the local analysis window. For each connecting edge belonging to the local analysis window, an angle array of the continuous change path of the fiber orientation angle corresponding to that connecting edge is obtained. Each angle sample value is extracted one by one from the angle array; each angle sample value is a directional angle. All directional angles corresponding to all connecting edges are summarized to form a directional angle set, and the number of elements in the directional angle set is denoted as . .

[0055] In practice, the circular variance of all directional angles is calculated, and this circular variance is used as the dispersion of the fiber orientation angles. The dispersion of the fiber orientation angles is denoted as... The first in the set of directions and angles Each direction angle is denoted as The unit of angle is degrees. The value ranges from 0 degrees to 180 degrees. Considering that fiber orientation angles have no directional distinction, each direction angle is... Multiply by 2 to convert to an angle value within the range of 0 to 360 degrees, denoted as , The formula for calculating the circular variance uses a statistic based on the length of the resultant vector, and is expressed as follows:

[0056]

[0057] in, This indicates the dispersion of fiber orientation angles. The value ranges from 0 to 1, when all directions are exactly the same angle. The value is 0 when the direction angle is uniformly distributed in all directions. The value is close to 1; This represents the total number of direction angles in the set of direction angles. The range of values ​​for is positive integers; Indicates the first The angle value is obtained by multiplying the angle in each direction by 2, and the unit is degrees. The value ranges from 0 degrees to 360 degrees; express The value of the sine function; express The value of the cosine function; Indicates will indivual The sine values ​​of the angles are summed. Indicates will indivual The cosine values ​​of the angles are summed. It represents the length of the resultant vector of the set of directions and angles.

[0058] In practical implementation, the degree of aggregation of key nodes will be considered. and the dispersion of fiber orientation angle These are combined into a two-dimensional vector, which serves as the local classification feature vector. The local classification feature vector is denoted as... Its form is The first component represents the clustering degree of key nodes, and the second component represents the dispersion of fiber orientation angles. A corresponding local classification feature vector is calculated for each local analysis window, and the local classification feature vectors of all local analysis windows constitute a feature vector set.

[0059] See Figure 6 The horizontal axis in the figure represents the degree of clustering of key nodes within the local analysis window. The unit is per cubic micrometer, and the vertical axis represents the dispersion of the fiber orientation angle. The value ranges from 0 to 1. In the figure, circles (○) indicate the local classification feature vectors of high-density class samples, squares (□) indicate the local classification feature vectors of medium-density class samples, and triangles (△) indicate the local classification feature vectors of low-density class samples.

[0060] The distribution trend in the graph shows the degree of clustering of key nodes. The density categories show a clear increasing trend from low to high. High-density category feature points... The values ​​are mainly concentrated between 0.6 and 1.2 micrometers per cubic meter, in the medium-density category. The values ​​are mostly distributed between 0.3 and 0.8, with low-density categories. The values ​​are generally below 0.4. The dispersion of fiber orientation angles. Presentation and The opposite trend is observed in high-density categories. The values ​​are mainly concentrated between 0 and 0.4, indicating that the fiber orientation is relatively consistent; medium density category The values ​​range from 0.4 to 0.7, with moderate directional dispersion; low-density category The values ​​are mostly concentrated in the range of 0.7 to 1.0, indicating that the fiber orientation is highly discrete.

[0061] Furthermore, the local classification feature vectors of the three density categories in the figure exhibit good discriminative power in the two-dimensional feature space. Each of the three categories forms a relatively concentrated clustered distribution, with clear boundary separation between the clusters. This distribution characteristic provides an effective basis for training the support vector machine model based on the combination of key node aggregation degree and fiber orientation angle dispersion in Example 3, thereby supporting accurate classification of high, medium, and low density categories. Overall, the data in the figure intuitively reflects the important role of the spatial aggregation of key nodes and fiber orientation consistency within the density feature backbone network in the density classification of hydrophilic fiber dressings.

[0062] Example 4:

[0063] In the specific implementation, multiple hydrophilic fiber dressing samples of different density categories were collected. These density categories included high-density, medium-density, and low-density categories, with an equal number of samples collected for each category. For each hydrophilic fiber dressing sample, following a complete processing flow—acquiring multi-scale microscopic image sequences, performing fiber trajectory tracking, constructing a cross-layer fiber network topology, generating a density feature backbone network, dividing into multiple local analysis windows, and calculating the local classification feature vector for each local analysis window—the local classification feature vectors for all local analysis windows of the hydrophilic fiber dressing sample were extracted. Each local analysis window's local classification feature vector is a two-dimensional vector, with the first component representing the clustering degree of key nodes. The second component is the dispersion of the fiber orientation angle. For each local classification feature vector, label its corresponding overall density category of the hydrophilic fiber dressing sample. The labeling format is a category label value: high density category is labeled with category label value 1, medium density category with category label value 2, and low density category with category label value 3. Combine the local classification feature vectors and category label values ​​corresponding to all local analysis windows of all hydrophilic fiber dressing samples to construct a training dataset. Each sample in the training dataset consists of a two-dimensional feature vector and a category label value.

[0064] In practical implementation, the Support Vector Machine (SVM) model is used as the basic framework for the window-level classifier. The SVM model employs a non-linear classifier structure, with radial basis functions (RBFs) as the kernel function. The mathematical form of the RBF is: .in, This represents the result of the kernel function operation. and Let each represent any two distinct local classification feature vectors in the training dataset. This represents the Euclidean distance between two local classification feature vectors. The kernel width parameter of the radial basis functions. The range of values ​​for is positive real numbers. The value is set to 0.5, based on cross-validation of the validation dataset within a parameter range of 0.01 to 2.0 using a grid search method. When the threshold is 0.5, the classification accuracy on the validation dataset reaches its maximum value. The decision function of the support vector machine model adopts a one-to-one multi-class strategy, training three binary sub-support vector machines by pairwise combinations of the high-density, medium-density, and low-density classes. The training process of each binary sub-support vector machine involves finding the maximum margin hyperplane between the corresponding two classes of training samples. The mathematical form of the maximum margin hyperplane satisfies the constraints. and Simultaneously minimize the objective function .in Indicates the first in the training dataset A local classification feature vector, Indicates the first The temporary binary classification label value of each local classification feature vector in the current binary classification problem. The temporary binary classification label value takes the value of +1 or -1. Describes the normal vector of the hyperplane. The bias term represents the hyperplane. Indicates the first Slack variables for each sample, This represents the penalty coefficient. The value is set to 1.0, based on the principle of evaluating the balance between overfitting and underfitting on the training dataset during training. When the value is 1.0, the model has the best generalization ability on unseen data.

[0065] In practice, the local classification feature vector of each local analysis window is used as the input feature, and the overall density class corresponding to the local analysis window is used as the supervision label to train the support vector machine model. The training process uses a sequential minimum optimization algorithm to iteratively solve the Lagrange multipliers of the three binary sub-support vector machines. The convergence condition is that all Lagrange multipliers satisfy the Carlow-Kun-Tucker condition within a tolerance range, with the tolerance value set to 10 to the power of -3. After training, for any input local classification feature vector, the three binary sub-support vector machines output confidence scores indicating whether the local classification feature vector belongs to the two corresponding density classes. A voting method is used to merge the classification results of the three binary sub-support vector machines, and the number of votes obtained for each density class is normalized to a probability value. The normalization method is to add a smoothing factor to the number of votes for each density category and then divide by the sum of the total number of votes and three times the smoothing factor. The smoothing factor is set to 1.0. After normalization, the probability values ​​of the local analysis window belonging to the high-density category, the medium-density category, and the low-density category are obtained. The three probability values ​​constitute the window category probability distribution.

[0066] In practice, the local classification feature vector of each local analysis window of the hydrophilic fiber dressing sample to be classified is input into the trained support vector machine model. Following the complete processing flow, the local classification feature vectors of all local analysis windows of the current hydrophilic fiber dressing sample to be classified are extracted. Each local classification feature vector is then fed into the three binary sub-support vector machines of the support vector machine model to calculate the voting score, and after normalization, the window category probability distribution corresponding to each local analysis window is obtained.

[0067] In practice, the set of adjacent windows for each local analysis window is determined based on the spatial position of its center point in the three-dimensional coordinate system of the density feature backbone network. The X-coordinate of the center point of a local analysis window is equal to the midpoint of its boundary range along the X-axis, the Y-coordinate is equal to the midpoint of its boundary range along the Y-axis, and the Z-coordinate is equal to the midpoint of its boundary range along the Z-axis. For each local analysis window, all other local analysis windows are traversed, and the three-dimensional Euclidean distance between the center points of two local analysis windows is calculated. The three-dimensional Euclidean distance is calculated as the square root of the sum of the squares of the differences in the X, Y, and Z coordinates of the two center points. When the calculated three-dimensional Euclidean distance is less than the size of one window, the corresponding other local analysis windows are added to the set of adjacent windows of that local analysis window. A window size is defined as the magnitude of the vector formed by the side lengths of a three-dimensional sliding window in the X-axis direction, the Y-axis direction, and the Z-axis direction.

[0068] In practice, for each density category, a weighted average is taken between the corresponding probability value in the window category probability of the local analysis window and the corresponding probability values ​​in the window category probabilities of all neighboring windows. The formula for calculating the weighted voting score of a certain density category in this local analysis window is:

[0069]

[0070] in, Indicate density category The weighted voting score in this local analysis window, The value of is a real number ranging from 0 to 1; Indicates density category identifier, A value of 1 indicates a high-density category. A value of 2 indicates a medium density category. A value of 3 indicates a low-density category; This indicates that the window category probability of the local analysis window itself belongs to the density category. The probability value; This represents the weighting coefficient of the local analysis window itself; This represents the total number of adjacent windows in the adjacent window set; Represents the first in the set of adjacent windows Among the window category probabilities of adjacent windows, the one belonging to the density category is... The probability value; Represents the first in the set of adjacent windows The weighting coefficients of each adjacent window. and The weighting factor is determined by taking the reciprocal of the distance between adjacent windows. The distance between adjacent windows is defined as the three-dimensional Euclidean distance between the center point of the local analysis window and the center point of the corresponding adjacent window in the set of adjacent windows. When the distance between adjacent windows is zero, the weighting factor takes a preset maximum value. Adaptive adjustment is then performed by combining the distance from the center point of the local analysis window to the geometric center of the hydrophilic fiber dressing sample. Specifically, the distance from the center point of the local analysis window to the geometric center of the hydrophilic fiber dressing sample is calculated. The geometric center of the hydrophilic fiber dressing sample is defined as the location point corresponding to the average spatial distribution coordinates of all key nodes in the density feature backbone network. This distance value is mapped to an adjustment factor. The adjustment factor is calculated as the negative exponential function value of the ratio of the distance value to the maximum radius of the coverage area of ​​the density feature backbone network in three-dimensional space. The adjustment factor is then multiplied by the original weighting factor to obtain the adaptively adjusted weighting factor.

[0071] In practice, the weighted voting scores of all local analysis windows across the three density categories are used. The scores are summed separately to obtain the cumulative scores for the high-density category, the medium-density category, and the low-density category. The three cumulative scores are compared, and the density category with the highest cumulative score is selected as the overall density category label for the hydrophilic fiber dressing sample to be classified. The overall density category label is output as one of the following: high-density category label, medium-density category label, or low-density category label.

[0072] Example 5:

[0073] In specific implementation, please refer to Figure 3 The hydrophilic fiber dressing sample is fixed on the stage of the microscopic imaging platform. The stage surface of the platform has a sample positioning groove, the depth of which is half the thickness of the hydrophilic fiber dressing sample. The length and width of the groove match the length and width of the sample, respectively. The hydrophilic fiber dressing sample is placed flat within the positioning groove, ensuring its lower surface is in close contact with the bottom of the groove and its upper surface is flush with the stage surface. After placement, the sample is pressed against its perimeter using elastic clamps along the stage edge, ensuring the pressed area does not cover the central imaging area.

[0074] In practice, a vertical scanning range is defined from the upper to the lower surface of the hydrophilic fiber dressing sample. Using the stage plane of the microscopic imaging platform as the reference zero plane for vertical positioning, the starting point of the vertical scanning range is set at a first distance above the stage plane. This first distance is equal to the nominal thickness of the hydrophilic fiber dressing sample plus an upward offset of 50 micrometers. This offset is set to compensate for the increase in thickness caused by moisture absorption and expansion of the hydrophilic fiber dressing sample, as well as errors in the stage plane's positioning. The ending point of the vertical scanning range is set at a second distance below the stage plane, with a value of 20 micrometers. This second distance is set to ensure that the imaging focal plane completely penetrates the lower surface of the hydrophilic fiber dressing sample, covering any possible gap between the hydrophilic fiber dressing sample and the bottom surface of the sample positioning groove. The total length of the vertical scanning range is equal to the vertical distance between the starting and ending points.

[0075] In practice, the vertical scanning range is divided into multiple imaging layers according to a fixed interlayer spacing. The value of the fixed interlayer spacing is determined based on the depth of field of the objective lens used in the microscopic imaging system, and is set to two-thirds of the objective lens's depth of field. The objective lens depth of field is calculated using the standard depth of field formula in microscopic optics systems, and the depth of field depends on the numerical aperture of the objective lens and the center wavelength of the light source used for imaging. When the center wavelength of the light source used for imaging is 550 nm and the numerical aperture of the objective lens is 0.3, the calculated objective lens depth of field is 6.1 μm, and the fixed interlayer spacing, after being set to two-thirds of the objective lens depth of field, is 4.0 μm. Starting from the beginning of the vertical scanning range, an imaging layer is set at intervals of one fixed interlayer spacing. The depth position of the imaging layer is represented by the vertical distance from the center plane of the imaging layer to the stage plane. The depth positions of all imaging layers form an arithmetic sequence, with the first term being the starting depth value of the vertical scanning range, the common factor being the negative of the fixed interlayer spacing, and the last term being the depth value of the last imaging layer in the vertical scanning range whose depth value does not exceed the end point of the vertical scanning range.

[0076] In practice, at each imaging layer, the autofocus system adjusts the focal length to the depth of that layer. The autofocus system includes a focusing motor, a grating ruler displacement sensor, and a feedback controller. For each imaging layer, the target displacement of the focusing motor is calculated based on the depth of that layer. The formula for calculating the target displacement is:

[0077]

[0078] in, This indicates the target displacement of the focusing motor, in micrometers. The value range is within the mechanical stroke range of the focusing motor; This represents the vertical distance from the center plane of the imaging layer to the stage plane, expressed in micrometers. The value decreases sequentially from the start to the end of the vertical scan range; This represents the calibration offset between the objective lens focal plane and the stage plane during the initial focusing of the imaging system, expressed in micrometers. The value is obtained through automatic focusing calibration on a standard scale target. The target displacement is converted into the number of drive pulses for the focusing motor and sent to the focusing motor driver, which then drives the objective lens to move vertically. The grating ruler displacement sensor detects the actual displacement of the objective lens in real time and feeds it back to the feedback controller. The feedback controller compares the actual displacement with the target displacement. When the deviation between the actual and target displacements is less than a preset positioning tolerance, the feedback controller outputs a hold signal to lock the objective lens position. The positioning tolerance is set to 0.2 micrometers, based on the allowable range of the objective lens depth of field for image sharpness.

[0079] In practice, after the autofocus system completes the imaging focal length adjustment, the camera is triggered to acquire the original microscopic image corresponding to the imaging layer. The camera is a monochrome scientific-grade complementary metal-oxide-semiconductor (CMOS) camera with a pixel resolution of 2048 pixels by 2048 pixels and a pixel depth of 12 bits. The camera's exposure time is automatically adjusted based on the average reflectance of the hydrophilic fiber dressing sample. The automatic adjustment strategy involves metering the pre-scanned area of ​​the hydrophilic fiber dressing sample before imaging layer acquisition, with the exposure time adjustment target being that the 95th percentile of the pixel brightness histogram in the metering result reaches 80% of the saturation value. Simultaneously with triggering the camera to acquire the original microscopic image, the depth coordinates of the original microscopic image are recorded. The depth coordinates are equal to the vertical distance from the center plane of the imaging layer to the stage plane, with a recording accuracy of 0.1 micrometers. The depth coordinates are stored in a one-to-one correspondence with the original microscopic image, with the storage format being that the depth coordinates are written into the metadata tag of the original microscopic image.

[0080] In practice, all acquired raw microscopic images are arranged into a sequence according to their depth coordinate values, from smallest to largest. The ascending depth coordinate values ​​correspond to the spatial order of the imaging layers, from the top surface to the bottom of the hydrophilic fiber dressing sample. Illumination non-uniformity correction is performed on each raw microscopic image in the sequence. This correction employs a flat-field correction method based on a reference white field image. Before acquiring the raw microscopic images of the hydrophilic fiber dressing sample, a standard reflectance white plate is imaged using the same microscopic imaging system under the same illumination conditions and objective magnification, resulting in multiple reference white field images and a calculated average white field image. For each raw microscopic image in the sequence, the gray value of each pixel in the raw image is divided by the corresponding pixel's gray value in the average white field image, and then multiplied by the average gray value of all pixels in the average white field image to obtain the image after illumination non-uniformity correction. Illumination non-uniformity correction eliminates brightness deviations caused by uneven intensity distribution of the illumination source in the microscopic imaging system and vignetting effects of the optical system.

[0081] In practice, each original microscopic image in the sequence undergoes noise filtering preprocessing after illumination inhomogeneity correction. The noise filtering preprocessing employs a nonlocal means denoising algorithm. The search window size for the nonlocal means denoising algorithm is set to 7 pixels by 7 pixels, the similarity block size is set to 3 pixels by 3 pixels, and the attenuation coefficient is calculated based on the estimated noise standard deviation of the original microscopic images. The estimated noise standard deviation is obtained by statistically analyzing the pixel grayscale values ​​of the uniform background region in the original microscopic images. The nonlocal means denoising algorithm removes shot noise and readout noise introduced by the complementary metal-oxide-semiconductor camera while preserving the detailed structure of fiber edges and fiber cross nodes. All the original microscopic images after illumination inhomogeneity correction and noise filtering preprocessing are combined in ascending order of depth coordinate values ​​to form a multi-scale microscopic image sequence.

[0082] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for classifying the image features of hydrophilic fiber dressing density, characterized in that, The method includes: A multi-scale microscopic image sequence is obtained by performing layered scanning of a hydrophilic fiber dressing sample by changing the imaging focal length. A fiber trajectory tracing operation is performed on the multi-scale microscopic image sequence to extract the spatial distribution coordinates of fiber cross nodes and the continuous change path of fiber orientation angle in each layer of microscopic image; Based on the spatial distribution coordinates of the fiber cross nodes, a cross-layer fiber network topology is constructed, and key nodes traversed by the continuously changing paths of the fiber orientation angle are marked on the fiber network topology to generate a density feature backbone network. The density feature backbone is divided into multiple local analysis windows, and the degree of clustering of key nodes and the dispersion of fiber orientation angle within each local analysis window are calculated to obtain local classification feature vectors. The local classification feature vectors of all local analysis windows are weighted according to spatial adjacency to generate a weighted vote, and the overall density category label of the hydrophilic fiber dressing sample is output.

2. The image feature classification method for the density of hydrophilic fiber dressings according to claim 1, characterized in that, Perform fiber trajectory tracing on the multi-scale microscopic image sequence to extract the continuous variation path of the spatial distribution coordinates of fiber intersection nodes and fiber orientation angles in each layer of microscopic image, including: Gradient orientation histogram matching is applied to the current layer microscopic image in the multi-scale microscopic image sequence to identify the fiber pixel point set, and pixels whose curvature changes exceed a set threshold are selected from the fiber pixel point set as candidate intersection points. Candidate intersection points with at least three fiber extension directions within their eight neighborhoods are identified as fiber intersection nodes, and the positions of the fiber intersection nodes in the image coordinate system are recorded as the spatial distribution coordinates. The fiber segment between two adjacent fiber intersection nodes is refined by skeleton line refinement. The tangent direction is calculated pixel by pixel along the skeleton line, and the angle sequence of all tangent directions is used as the continuous change path of the fiber orientation angle.

3. The image feature classification method for the density of hydrophilic fiber dressings according to claim 2, characterized in that, The process of constructing a cross-layer fiber network topology map based on the spatial distribution coordinates of the fiber intersection nodes, and marking key nodes traversed by continuously changing paths of fiber orientation angles on the fiber network topology map to generate a density feature backbone network includes: The spatial distribution coordinates of fiber cross nodes extracted from each layer of microscopic image are mapped to a unified three-dimensional spatial coordinate system. The node connection relationship between layers is established by comparing the angle between the end direction and the starting direction of the continuous change path of fiber orientation angle between adjacent microscopic images of different layers. Using the fiber cross nodes as vertices and the node connection relationships as edges, a cross-layer fiber network topology graph is generated. Traverse each edge in the cross-layer fiber network topology graph, and determine whether there is a sudden change point where the angle jump exceeds the jump threshold on the continuous change path of the fiber orientation angle between the two fiber intersection nodes corresponding to the edge. If there is, mark the fiber intersection node corresponding to the sudden change point as the key node. The density feature backbone network is composed of all key nodes and their connecting edges.

4. The image feature classification method for the density of hydrophilic fiber dressings according to claim 3, characterized in that, The fiber cross nodes corresponding to the mutation points where the angle jump exceeds the jump threshold are assigned higher topological connectivity weights in the density feature backbone network.

5. The image feature classification method for the density of hydrophilic fiber dressings according to claim 3, characterized in that, The process involves dividing the density feature backbone into multiple local analysis windows and calculating the clustering degree of key nodes and the dispersion of fiber orientation angles within each local analysis window to obtain local classification feature vectors, including: A three-dimensional sliding window of fixed size is set on the density feature backbone network, and the three-dimensional sliding window is moved according to a preset step size. The area covered by the window after each sliding is used as a local analysis window. For each local analysis window, the total number of key nodes contained within the window is counted, and the total number is divided by the volume of the window to obtain the degree of clustering of the key nodes; For each local analysis window, extract all directional angles on the continuous change path of the fiber orientation angle corresponding to all connecting edges within the window, calculate the circular variance of all directional angles, and use the circular variance as the dispersion of the fiber orientation angle. The degree of clustering of the key nodes and the dispersion of the fiber orientation angle are combined into a two-dimensional vector, which is used as the local classification feature vector.

6. The image feature classification method for the density of hydrophilic fiber dressings according to claim 5, characterized in that, The method involves weighting the local classification feature vectors of all local analysis windows according to their spatial adjacency to output the overall density category label of the hydrophilic fiber dressing sample, including: Based on the spatial position of the center point of each local analysis window in the density feature backbone, the set of neighboring windows of the local analysis window is determined, wherein the distance between each neighboring window in the set and the local analysis window in three-dimensional space is less than one window size; For each local classification feature vector of a local analysis window, a window-level classifier is pre-trained, and the window-level classifier outputs the probability distribution of the window belonging to different density categories. For each density category, the corresponding probability value in the window category probability of the local analysis window is weighted and averaged with the corresponding probability value in the window category probability of all neighboring windows. The weighting coefficient is inversely proportional to the distance between neighboring windows, thus obtaining the weighted voting score of the local analysis window. The weighted voting scores of all local analysis windows are summed up, and the density category with the highest summation score is selected as the overall density category label.

7. The image feature classification method for the density of hydrophilic fiber dressings according to claim 3, characterized in that, The process of mapping the spatial distribution coordinates of fiber intersection nodes extracted from each layer of microscopic images to a unified three-dimensional spatial coordinate system, and establishing inter-layer node connection relationships between adjacent microscopic images of different layers by comparing the angle between the endpoint direction and the starting direction of the continuous change path of fiber orientation angle, includes: Obtain the endpoint direction angle of the continuous change path of the fiber orientation angle corresponding to a certain fiber intersection node in the previous layer microscopic image, and the starting direction angle of the continuous change path of the fiber orientation angle corresponding to a certain fiber intersection node in the current layer microscopic image. Calculate the absolute difference between the endpoint direction angle and the starting point direction angle. When the absolute difference is less than the angle connectivity threshold, create an interlayer connection edge between the fiber cross node in the previous layer micrograph and the fiber cross node in the current layer micrograph. Repeat the above comparison process, traversing all fiber cross nodes in all adjacent layer micrographs, to complete the establishment of the node connection relationship between the layers.

8. The image feature classification method for the density of hydrophilic fiber dressings according to claim 3, characterized in that, The process involves traversing each edge in the cross-layer fiber network topology graph, determining whether there is a sudden change point where the angle jump exceeds a threshold on the continuous path of the fiber orientation angle between the two fiber intersection nodes corresponding to that edge. If such a point exists, the fiber intersection node corresponding to the sudden change point is marked as the key node. All key nodes and their connecting edges constitute the density feature backbone network, including: Select an edge from the cross-layer fiber network topology diagram, obtain the first fiber cross node and the second fiber cross node connected by the edge, and extract all angle sampling values ​​on the continuous change path of the fiber orientation angle from the first fiber cross node to the second fiber cross node. Calculate the jump amplitude between two adjacent angle sampling values. If there is a jump amplitude that exceeds the jump threshold, it is determined that there is a sudden change point on the continuous change path, and the first fiber cross node and the second fiber cross node are both marked as key nodes. After traversing all edges in the cross-layer fiber network topology, collect all fiber cross nodes marked as key nodes, and delete the fiber cross nodes and their connecting edges that are not marked as key nodes from the cross-layer fiber network topology to obtain the density feature backbone network.

9. The image feature classification method for the density of hydrophilic fiber dressings according to claim 6, characterized in that, For each local classification feature vector of a local analysis window, a window-level classifier is pre-trained. The window-level classifier outputs the probability distribution of the window belonging to different density categories, including: Collect multiple hydrophilic fiber dressing samples of different density categories, extract the local classification feature vectors of all local analysis windows of each sample according to the method described in claim 4, and label the overall density category of the sample to which each local classification feature vector belongs to, and construct a training dataset. The support vector machine model is used as the basic framework of the window-level classifier. The local classification feature vector of each local analysis window is used as the input feature, and the overall density class corresponding to the window is used as the supervision label. The support vector machine model is trained so that the trained support vector machine model can output the probability value of each local analysis window belonging to different density classes. The local classification feature vector of each local analysis window of the hydrophilic fiber dressing sample to be classified is input into the trained support vector machine model to obtain the window category probability distribution corresponding to each window.

10. The image feature classification method for the density of hydrophilic fiber dressings according to claim 1, characterized in that, During the weighted voting process, the weight coefficients of different local analysis windows are adaptively adjusted based on the distance from the center point of the window to the geometric center of the hydrophilic fiber dressing sample.

Images