Traditional Chinese medicinal material defect detection method based on machine vision

Abstract

The invention discloses a traditional Chinese medicinal material defect detection method based on machine vision, relates to the technical field of intelligent detection, and solves the problems that in the prior art, a threshold cannot be migrated, so that an inter-batch judgment standard is unstable, and high-resolution block reasoning causes boundary pseudo defects and instance wrong affiliation. According to the method, a reference area is constructed in a batch domain, a statistical descriptor is extracted, and monotone mapping calibration is performed on defect scores so as to realize cross-batch unified judgment; meanwhile, a seam continuity field is introduced in the image block splicing process, defect connected domains are fused through a cross-window adjacency relation, and accurate attribution of defects and medicinal material pieces is achieved based on bipartite graph matching; according to the method, the judgment stability and the statistical reliability in the traditional Chinese medicinal material defect detection process are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection technology, and more specifically to a method for detecting defects in Chinese medicinal materials based on machine vision. Background Technology

[0002] Chinese medicinal herbs are diverse and exhibit significant individual differences. Roots, tubers, root slices, and leaves naturally differ in morphology, cross-sectional structure, and surface color. Furthermore, factors such as processing techniques, harvesting period, origin (authenticity), moisture content, and slice thickness can cause systematic shifts in color, texture, and microstructure. Common appearance defects include, but are not limited to, surface mold spots, mold cores, insect holes, longitudinal and transverse cracks, mechanical abrasions, browning, foreign matter inclusions (mud, mixed roots), and uneven slices. Because defects range in size from tiny mold spots to large damaged slices and are often related to the histological characteristics of the medicinal material (such as vascular bundles, fiber arrangement, and xylem exposure), purely manual visual inspection is time-consuming and subject to subjective influences. Detection rates and consistency fluctuate significantly with the inspector's experience, visual fatigue, and inspection conditions, making it difficult to meet the requirements of high throughput and traceable release in online production lines.

[0003] The industry is gradually adopting machine vision for online inspection and batch release. Current practices typically use two-dimensional images as a basis, performing preprocessing and morphological / color / texture feature analysis, or outputting defect scores or quality grades after multi-view / temporal alignment. For example, CN119693343A proposes a comprehensive scoring and archiving method based on morphology, color, and texture, aiming to reduce human intervention and improve the method's versatility and accuracy; CN120374870A improves judgment stability by using timed acquisition, pixel alignment, and multi-view geometric constraints to suppress offsets caused by lighting and jitter. While these existing methods are effective in typical scenarios, in actual production, whether it's single-image scoring based on color / texture / morphology or similarity / classification scores obtained through alignment and multi-view fusion, the final decision for release must fall within a fixed or limited threshold. The inter-batch distribution drift of Chinese medicinal materials (origin, harvesting season, moisture content, cleanliness, cutting thickness, and processing differences) can cause the overall score to shift or shrink, resulting in missed or over-detection of the same threshold in different batches. This affects the consistency of AQL control and release, forcing inspection and maintenance personnel to frequently adjust thresholds or rebuild reference sets, thus reducing the cross-batch comparability of quality traceability.

[0004] Furthermore, to balance field of view and resolution, production lines often divide large images into several tiles and stitch them together after reasoning about each tile. Medicinal herb slices are often slightly stacked or irregularly shaped, causing context truncation at the seams: tiny mold spots or fine cracks are more easily misjudged or fragmented at tile edges, causing "tears" in the semantic segmentation mask at the stitching points. Even slight deviations in post-processing merging can merge two adjacent small defects into a single large defect or lead to duplicate counting. In stacked structures, upper-layer defects are mistakenly recorded as lower-layer good products, and simple statistics based on defect percentage amplify the defect rate. Even with multi-view geometric constraints to suppress biases caused by lighting and jitter, images still require block segmentation, cross-view fusion, and instance alignment. Instance merging at seams, cross-window adjacency, and defect attribution still rely on post-processing strategies and heuristic rules. Different stitching strategies or merging thresholds introduce systematic statistical biases, further amplifying FAR / FRR jumps and batch-to-batch distortions caused by threshold instability. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention discloses a machine vision-based method for detecting defects in Chinese medicinal materials. The method aims to maintain the stability of the judgment threshold and the consistency of defect statistics under complex batch conditions and high-resolution block detection scenarios, thereby overcoming the problems of fluctuation in detection standards and distortion of statistical caliber in existing technologies.

[0006] To achieve the above-mentioned technical effects, the present invention adopts the following technical solution: A machine vision-based method for detecting defects in traditional Chinese medicinal materials includes: Step 1: Acquire continuous frame images of Chinese medicinal materials, generate a set of sub-windows according to a fixed window and overlap width, and establish a seam index; Step 2: Within the sub-window set, filter the anchor region set based on gradient variance, texture energy, and color entropy, and remove the seam neighborhood. Step 3: Extract color histogram statistics, power spectral moments, and gray-level co-occurrence moments from the anchor region set, construct a batch domain descriptor, and read the reference descriptor and reference quantile group; Step 4: Score all sub-windows for defects, obtain the original score set, and count the reference area quantiles as the calibration benchmark. Step 5: Based on the difference between the batch domain descriptor and the reference descriptor, construct a monotonic segmented mapping for the fraction domain, set quantile constraints, obtain mapping parameters, calibrate the original fractions, and set a fixed threshold in the calibration domain. Step 6: Calculate the structural tensor direction and the difference in strength across the joint on both sides of the joint strip, generate the joint continuity field and establish a boundary mask; Step 7: Perform connected component labeling based on calibration score and threshold within each sub-window and map it to global coordinates to construct a cross-window adjacency graph composed of seam continuity and shape compatibility, and generate a global defect instance set through merging and nonmaximum suppression. Step 8: Generate the foreground response map of the medicinal material slices, obtain the slice instance set through distance transformation and watershed method, and repair the seam contour according to the seam continuity field cost; Step 9: Construct a bipartite graph of defect instances and sheet instances. The edge weights are given by the region intersection-union ratio, boundary direction difference, and seam continuity integral. Solve for the maximum weight matching to obtain the defect attribution mapping. Step 10: Generate defect statistical labels based on the defect attribution mapping, write the batch domain descriptor, mapping parameters, reference quantile group, judgment threshold and seam continuity summary into the batch file, and monitor drift trigger parameter updates within the sliding time window.

[0007] Based on the above technical solution, the positive and beneficial effects of the present invention are as follows: This invention constructs a reference region in the batch domain and extracts batch descriptors to calibrate the original defect scores by establishing a segmented monotonic mapping. This ensures that the scores of different batches are unified into a stable reference system before entering the judgment stage. This avoids the situation in the prior art where the correspondence between scores and thresholds drifts as a whole with batch differences such as origin, season, and dryness. It also reduces over-rejection or under-rejection caused by non-transferable thresholds and ensures the comparability of cross-batch detection results.

[0008] This invention introduces a seam continuity field during image segmentation reasoning and result stitching, and establishes an adjacency graph based on directional consistency and seam difference to perform defect connectivity fusion, thereby suppressing false defects in the boundary region and ensuring that the defect mask remains coherent at the stitching point, avoiding the situation where the same defect is counted repeatedly or adjacent small defects are mistakenly merged; at the same time, by matching the defect with the bipartite graph of the sheet, the problem of small defects being mistakenly assigned to the lower layer of good products in stacked samples is solved. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 This is a flowchart of the steps of the present invention; Figure 2 A schematic diagram illustrating the construction of the batch field descriptor of this invention; Figure 3 This is a schematic diagram illustrating the foreground response and watershed correction of the present invention; Figure 4 This is a schematic diagram of the seam tape of the present invention. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0011] To facilitate understanding of this invention, a machine vision-based method for detecting defects in traditional Chinese medicinal materials disclosed in this application is described in detail below. Please refer to [link to relevant documentation]. Figure 1 The diagram illustrates the steps of a machine vision-based method for detecting defects in Chinese medicinal materials. The method includes the following steps: Step 1: Acquire continuous frame images of Chinese medicinal herbs, generate a set of sub-windows according to a fixed window and overlap width, and establish a seam index; In this step, the acquisition module continuously acquires frames of medicinal herbs on the moving carrier at a resolution of not less than 2048×2048 pixels and a frame rate of not less than 15 frames / second at a fixed imaging calibration position. Please refer to [link to relevant documentation]. Figure 4 The schematic diagram of the seam tape shown is denoted as the grayscale or color image obtained from the acquisition. Where t is the time frame index, and x, y are the pixel coordinates; the camera's extrinsic and intrinsic parameters are calibrated to obtain the homography matrix H or intrinsic parameter matrix K, which serves as the basis for subsequent global coordinate mapping. The pixel-to-physical coordinate transformation is performed according to... Completed. The meanings of all symbols in the formula will be explained uniformly below. A set of sub-windows is generated based on the acquired images, with equal row and column spacing. The sub-window size... The recommended value range is 64×64 to 512×512 pixels, with a sub-window overlap rate r ∈ [0.10, 0.40] and a step size. The number of windows is calculated both horizontally and vertically. Each child window is uniquely identified by an ID (i,j,t), which records its top-left pixel coordinates and the corresponding global coordinate window range. The seam index is based on the length of the child window's shorter side. Determine bandwidth Where α ranges from 0.05 to 0.20, the seam strip is extended along the short side, and the cross-seam sub-windows are recorded as entries using an adjacency list structure { Relative offset The initial value of the confidence level c is determined by the inter-frame structural similarity. The estimation is performed together with the gradient distribution consistency. For dynamic acquisition scenarios, the conditions for triggering the reconstruction sub-window set and seam index when adding a new batch or changing the type of medicinal material are: batch switching signal, and the change in average brightness of consecutive frames exceeding the threshold. Or the inter-frame structural similarity is below a threshold Image coordinates to global coordinates mapping uses online RANSAC estimation of four-point singularity updates when perturbation or camera micro-movement is present. If the estimated inlier ratio is lower than... If a warning is triggered, the system will be recalibrated. To facilitate subsequent calculations of cross-window adjacency graphs and seam continuity integrals, the sub-window index is also stored in a local fast-access index file. Its fields include sub-window ID, pixel range, physical coordinate range, seam band identifier, list of adjacent sub-windows, and acquisition time window. Specifically, it can be stored in a key-value database or binary serialization format to meet real-time access requirements.

[0012] It should be noted that, in this application, a sub-window refers to a discrete sampling unit pre-gridized across the entire frame with a fixed size and overlap rate, which facilitates the establishment of stable cross-frame and cross-window adjacency relationships; the seam band refers to a pixel band of a certain width extending along the short side of the sub-window boundary and serves as the spatial domain for cross-window continuity analysis; and a batch refers to a collection of images acquired within the same production batch under different lighting, variety, and transportation conditions. Specific parameters such as resolution and frame rate can be determined based on the production line speed, camera performance, and type of medicinal materials, and are not strictly limited in this regard.

[0013] Step 2: Within the sub-window set, filter the anchor region set based on gradient variance, texture energy, and color entropy, eliminating seam neighborhoods. In practice, Step 2 first performs local normalization on each sub-window; the normalization result is called the fundamental matrix and serves as the input for subsequent index calculations. Gradient variance is calculated by first calculating the gradient magnitude. Obtain regional indicators ,in The response is either a Sobel or Scharr operator; texture energy is obtained through a set of Gabor filters or a multi-band fast Fourier transform, where the center frequency and direction are set according to common textures on the surface of medicinal materials; color entropy is defined according to Shannon. For the probability distribution of the K bins in the color histogram, K is recommended to be 16-64. Calculate the quantile distribution and take the quantile points for the above three types of indicators within the current sub-window set or the current batch. As an adaptive threshold, it is recommended to select 75% or 90% of the quantile to adjust the sensitivity.

[0014] Candidate region extraction first generates a preliminary saliency map based on pixel-level local thresholding and then extracts connected components. For each connected component, an R-test is performed as a joint criterion. and And require minimum boundary distance Area constraints It represents the distance from the nearest pixel in the connected region to the boundary or seam of the sub-window. 5-30 pixels is recommended. and Based on the particle size of the medicinal materials, candidates located within the seam zone index coverage area were all eliminated. Hierarchical connectivity analysis was performed on the retained candidates, employing a multi-threshold, tree-like merging strategy. Connectivity segmentation is performed at each intensity threshold, and overlapping nodes are merged according to IoU, generating a primary anchor region set. Cross-frame consistency detection is then performed on the primary anchor region set, and temporal correlation is calculated. Average and require Or Number of occurrences within consecutive frames suggestion suggestion Recommendation 2: Items not meeting this requirement are considered transient noise and discarded. Finally, spatial uniformity correction is performed on the selected anchor regions using morphological hole filling, merging of small connected patches, and region boundary smoothing (e.g., curvature-based multiple sampling and fitting). An attribute vector is output for each anchor region, including region mask, centroid coordinates, area, principal direction, etc. Along with the timestamp, it is written to a local index for subsequent steps. Texture energy refers to the cumulative energy spectrum across frequency bands, reflecting the fiber direction and periodic texture characteristics of the medicinal material; seam neighborhood refers to the pixel band covered by the seam index established in step 1, which becomes the candidate rejection area. Specific parameters can be adjusted according to camera resolution, transport speed, and medicinal material variety.

[0015] Step 3: Extract color histogram statistics, power spectral moments, and gray-level co-occurrence moments from the anchor region set, construct batch domain descriptors, and read reference descriptors and reference quantile groups. Specifically, the process of step 3 includes: processing the anchor region set output in step 2 one by one; first, transforming each anchor region image block from camera space to CIELab space through linear color space transformation; quantizing the three channels into 32 levels at preset intervals and performing L1 normalized histogram statistics; and then concatenating the normalized histogram array in channel order to form color histogram subvectors. Grayscale patch for anchor area Applying a two-dimensional fast Fourier transform, the amplitude spectrum With frequency domain center For the path center pair Perform radial piecewise integration, and define the first and second order power spectral moments as follows: ,Will After normalization, the power spectral vector is formed. For grayscale patches, press Constructing a gray-level co-occurrence matrix using linear quantization Take pixel spacing With direction Normalize each matrix into a probability matrix. Calculate the four types of statistics: Contrast Correlation ,energy Uniformity .

[0016] Four sets of quantities in four directions are concatenated in a fixed order to form a GLCM vector. Color histogram (H), power spectrum Concatenate them sequentially into a single-anchor description vector. After performing column centering and variance normalization on all anchor vectors in the current batch, principal component analysis is used to reduce the dimensionality to... Dimension, as a processing step for batch internal stability and computational reversibility, This can be set to maintain 95% variance or fixed at 64, whichever you determine best. Then, based on the anchor area... Calculate the area quantile By interval The anchors were assigned to the four quantile groups. Within-group statistic is the mean. With covariance .

[0017] If the sample size of a certain quantile Less than the threshold Then, according to the time sliding window, anchor samples are accumulated in the most recent K frames until... Alternatively, guided contraction covariance estimation can be used. To obtain a numerically stable inverse matrix, Estimate according to the Ledoit–Wolf rule. Divide each quantile group. This serves as the reference descriptor for this batch and is written into the batch file; simultaneously, the historical reference descriptor set is read, and inter-group difference measures are calculated. This information is then returned to the subsequent score domain calibration module as the assigned weight. The trigger condition is defined as: when the total number of anchors... or any quantile When this occurs, a time accumulation mechanism or a dimensionality reduction threshold adjustment strategy is triggered; when compared with the maximum historical reference... Exceeding the threshold At that time, the batch was labeled as "domain drift" for use in subsequent steps to re-evaluate the fractional domain mapping parameters.

[0018] It should be noted that, in this application, the batch domain descriptor refers to a statistical representation of several anchor sets, used to capture the overall apparent differences of the current batch of materials.

[0019] In one implementation of this invention, multiple Chinese medicinal materials were collected for defect assessment, and a constraint model was constructed that reflects the intrinsic physicochemical properties of the medicinal materials based on their image features. This model is based on batch-domain feature constraints using the content of ginsenosides Re, Rg1, and Rb in ginseng (cultivated ginseng) as a reference. The ginseng shape is selected as follows: the main root is spindle-shaped or cylindrical and straight, with a grayish-yellow surface. The upper part or the entire root has sparse, shallow, discontinuous coarse transverse lines and obvious longitudinal wrinkles. The lower part has branch roots and numerous slender fibrous roots, often with inconspicuous small wart-like protrusions. The rhizome is slender, with a few short and thick ones, and sparse or dense, deeply sunken stem scars in the upper and middle parts. The texture is relatively hard, the cross-section is pale yellowish-white, powdery, the cambium ring is brownish-yellow, and the cortex has yellowish-brown dotted resin ducts and radial fissures. Among them, 20 branches are plump and robust, with the largest diameter main root, shaped like a short, thick cylinder or cone. The rhizome is thick and long, with dense and clearly defined stem scars. The texture is tight and deep, with tough fibrous roots and obvious pearl-like nodules. The average diameter of the main root is >2.0 cm, with a small length-to-width ratio, making it appear robust. The epidermis is dark grayish-yellow with a stable color. 30 stems are symmetrical, with a main root diameter of 1.7-2.0 cm, a distinct rhizome, and clear stem scars. The texture is clear, and pearl-like nodules are visible on the fibrous roots. The color is normal, grayish-yellow. 40 stems are of moderate size, slightly thinner than 30 stems, but still relatively full overall, with a main root diameter of approximately 1.4-1.7 cm. More slender individuals begin to appear. The texture is clear, but the depth may be slightly shallower. The color is similar to 30 stems. 50-60 stems: Significantly slender, with a small main root diameter (<1.4 cm), possibly with more branches or a relatively high proportion of fibrous roots. The rhizome is relatively short and thin. Increased morphological diversity with relatively shallower texture and less pronounced iron wire patterns. Slightly lighter color. 70 branches: Slender build with a thin taproot. Commonly branchesed, with potentially numerous and disordered fibrous roots. Possibly relatively light in texture. Mostly irregular in shape, slender and small. Shallow and sparse texture. Light color and weak texture.

[0020] Analysis of physicochemical test results from samples from multiple sources revealed a clear statistical regularity between the number of ginseng (cultivated ginseng) pieces and slice diameter and the content of the main active ingredients. Based on continuous sampling experiments, the ginsenoside content in samples with different indices and slice diameters showed the following patterns: Ginsenoside Rg1 showed a pattern of "40 pieces > 20 pieces > 50 pieces > 30 pieces > 70 pieces > 60 pieces" and "0.6 slices > 1.6 slices > 1.0 slices > 1.4 slices > 0.8 slices". The sum of ginsenoside Rg1 and ginsenoside Re showed a pattern of "40 pieces > 50 pieces > 70 pieces > 20 pieces > 30 pieces > 60 pieces" and "0.6 slices > 1.0 slices > 1.0 slices > 0.8 slices". The data shows the distribution trend of ginsenoside Rb content as follows: "60 > 40 > 70 > 30 > 50 > 20" and "1.6 > 0.8 > 1.6 > 1.0 > 1.4".

[0021] In the implementation of this invention, in order to ensure that the image features of Chinese medicinal materials can accurately reflect their intrinsic physicochemical properties, a batch domain feature constraint model is established with the content of Astragalus membranaceus isoflavone glucoside and total flavonoids as references.

[0022] Both wild-simulated and cultivated Astragalus membranaceus are cylindrical, thicker at the top, and some have branches with irregular longitudinal wrinkles or grooves. They are hard and tough, not easily broken, with a highly fibrous and powdery cross-section. The bark is yellowish-white, the wood is pale yellow, and there are radial striations. The difference between wild-simulated and cultivated Astragalus membranaceus lies in the roughness of the epidermis, the toughness of the root bark, the presence of cracks in the cross-section of the bark, the yellowish wood core, the loose texture, and the decayed, dark brown, or hollow center of the cut surface of older roots, but the area of ​​the cavity is no more than one-third of the cross-section. Cultivated Astragalus membranaceus has a smooth epidermis, a more flexible root bark, a dense cross-section, a yellowish-white central wood core, and a firm texture. In the experiment, parallel determinations of the physicochemical properties of wild-simulated and cultivated Astragalus membranaceus samples were conducted. The experimental results are as follows: Sample type Diameter d (cm) Verbena isoflavone glucoside content (average) (%) Total flavonoid content Level Classification Wild-like Sample A ≥1.4 0.0956 6.2954 Special Class Wild-like Sample B ≥1.3 0.0922 5.8456 First Class Wild-like sample C ≥1.2 0.0981 5.8163 Wild-like sample D ≥1.1 0.0779 5.1314 Second Class Wild-like sample E ≥1.0 0.0664 5.1013 Cultivation Sample F ≥1.2 0.0333 4.6592 Selection of goods Cultivation Sample G ≥1.1 0.05367 4.5754 Cultivation Sample H ≥1.0 0.06164 4.1948 Mixed goods As shown in the table, the content of verrucoside glucoside and total flavonoids in the simulated wild samples were generally higher than those in the cultivated samples, and showed a positive correlation with the increase of diameter. Based on this characteristic, when constructing the batch domain descriptor for Chinese medicinal materials, the segmented intervals of color histogram statistics, power spectral moments, and gray-level co-occurrence moments were mapped to this physicochemical distribution. Specifically, after extracting the anchor region features, the system normalizes and maps the mean texture energy of each sample to its corresponding verrucoside glucoside content, and uses the total flavonoid content as the texture gradient constraint term, so that the distribution of visual feature space maintains a one-to-one correspondence with the physicochemical space.

[0023] Analysis of physicochemical test results from multiple sources revealed a clear statistical regularity between the diameter of Astragalus membranaceus and the content of its main active ingredients. Based on continuous sampling experiments, the diameter distribution of verbenafiloside glucoside and total flavonoids in the samples showed a trend of: "wild-simulated sample 1.4cm > wild-simulated sample 1.3cm > wild-simulated sample 1.2cm > wild-simulated sample 1.1cm > wild-simulated sample 1.0cm > cultivated sample 1.2cm > cultivated sample 1.1cm > cultivated sample 1.0cm."

[0024] The above patterns indicate that within a group of medicinal materials with a uniform macroscopic appearance, structural compactness and color distribution can reflect changes in the internal chemical composition. In the implementation design of this invention, the correspondence between ginseng and astragalus is used for adaptive correction of the algorithm's feature layer to achieve synchronous changes in physicochemical and visual characteristics.

[0025] During the descriptor training phase, the texture responses of samples of different grades were compared and analyzed. The surface texture ROIs of premium, first, and second-grade samples were selected for spectral decomposition. The results showed that the high-frequency power spectrum peak of premium samples was significantly higher than that of lower-grade samples, and the gray-level co-occurrence moments were more concentrated in the variance range of the energy channels. This indicates that the higher the content of the intrinsic active substances in the medicinal material, the higher the directional consistency and energy concentration of its microscopic surface texture. Based on this, this invention established a content-texture bivariate regression model. By dynamically adjusting the standardized parameters of the power spectrum moments, the visual response curve of the sample grade maintained a linear fit with the chemical content within a 0.95 confidence interval (R²). 2 (≥0.93), achieving collaborative modeling of physicochemical indicators and machine vision features. Step 4: Score all sub-windows for defects, obtain the original score set, and statistically analyze the reference area quantiles as the calibration benchmark; please refer to the schematic diagram of batch domain descriptor construction as shown in Figure 2. In specific implementation, each sub-window... The feature vector has been obtained from step 3. (Statistical analysis of color histograms, frequency band power spectral moments, and gray-level co-occurrence moments, etc., spliced ​​in a fixed order). Classified by sub-window area to which the quantile group belongs. (See step 3 for dividing the area into 25%, 50%, and 75%) Calculate the within-group sample mean vector. With covariance matrix When the number of samples in a group Less than the preset threshold At this time, triggering time-accumulated or contraction covariance estimation guarantees matrix invertibility. To obtain a sub-window difference measure, a weighted quadratic form is defined. ,in, For the first Sub-window feature vector; This is the vector of the mean values ​​of the area quantile group to which the sub-window belongs; The regularized covariance matrix is ​​defined as follows: ,in , The feature dimension, This is the contraction coefficient (which can be estimated using Ledoit-Wolf or set as an empirical value, such as 0.05–0.2). For an identity matrix, regularization ensures that the matrix condition number is controlled and avoids singularities. —A diagonal weight matrix used to weight frequency bands / features according to signal-to-noise ratio. Weights The construction is as follows: First, the "band-type" characteristics (energy of power spectrum in several frequency bands) are indexed by frequency band. Estimating signal-to-noise ratio The recommended estimation formula is: ,in In frequency band For energy measures (e.g., the integral of the squared amplitude spectrum), the variances or variance-class statistics of the numerator and denominator are calculated on the anchor region set and the non-anchor background set, respectively. To prevent zero items. Then... Normalization yields the frequency band weights Map the frequency band weights to the corresponding dimensions of the feature vectors (set the positions of the frequency band class dimensions to the corresponding dimensions). For non-band features (such as color histogram and GLCM), after generalization by applying an inverse ratio to the intra-class variance or setting a constant baseline weight, the following results are obtained: (Diagonal element). The weighting design aims to amplify the spectral characteristics with higher signal-to-noise ratio in the current batch and suppress the interference of noise-dominated frequency bands on the metric.

[0026] In obtaining Then, compute the spatial-temporal local "quantile rank". This reflects the rarity of the sub-window within its neighborhood. Defined as the four neighboring sub-windows of the same frame and the corresponding sub-windows in the preceding and following T frames after spatial alignment (in the registered coordinate system), calculate the weight for each neighboring sample j: ,in The distance between the centers of space The number of frames in the time interval. Define the spatial / temporal scale parameters (in engineering, these can be set as a certain ratio between the half-width of the sub-window and the frame rate window); then define the weighted quantile rank. ,in Since it is an indicator function, therefore This rank naturally suppresses instantaneous noise in both space and time, and ensures stronger neighbor influence through distance attenuation.

[0027] raw fractions Using an amplitude-rarity fusion approach: , Measure the degree (magnitude) of deviation from the within-group statistical model. Reflecting local rarity (quantile ratio), taking the square root of their product maintains numerical scale stability and balances deviation strength and local significance in statistical interpretation. If real distance is required, it can be replaced with... Variants, etc.; when implementing them, they should be uniformly selected within the system and recorded in the batch file to facilitate reproducibility.

[0028] In implementation: When The number of conditions exceeds the threshold or When, shrink the covariance or add a regularization term; when Extremely larger than other samples in the batch (exceeding) Trigger an anomaly alarm and record it; when the number of valid samples in the neighborhood is lower than 100%. hour, It degenerates into local empirical quantiles or adopts global quantile compensation.

[0029] quantiles in the reference region From a subset of anchor regions with high cross-frame persistence and seamless neighborhood (e.g., average IoU in the most recent K frames) Extract the original score set from the anchor. The required quantile (e.g., the 90th or 95th percentile) is calculated as the calibration baseline and written to the batch file. Triggering conditions include: when the total number of anchors in the batch... If the sample size of any quantile group is insufficient, a time window accumulation strategy can be initiated or the confidence level of the quantile can be reduced to avoid unstable calibration.

[0030] It should be noted that in this application, "weighted Mahalanobis distance" specifically refers to a quadratic measure after applying a signal-to-noise ratio driven diagonal weight matrix W to the feature dimension, thereby enhancing the spectrum class discrimination capability.

[0031] Step 5: Based on the difference between the batch domain descriptor and the reference descriptor, construct a monotonic piecewise mapping for the fractional domain, set quantile constraints, obtain mapping parameters, calibrate the original scores, and set a fixed threshold in the calibration domain; specifically, first define batch difference quantization. Assume the batch domain descriptor is numbered according to the area quantile group. The current batch is in the The group's description vector is The historical reference descriptor is the mean vector. With covariance matrix Calculate the Mahalanobis distance scalar for each group. Constructing a difference vector .when or any If the mapping parameters are re-evaluated, the previous batch of mapping parameters will be used; otherwise, the trigger threshold will be used. It is set by engineering requirements.

[0032] The mapping function is implemented using a piecewise cubic polynomial under monotonic constraints, and the initial values ​​of the node evaluation points are set as several quantiles of the original fractions in the anchor region. , corresponding target value Calculated from the reference quantile group mapping target. Within the interval Above, define a standardized parameter for any input score s. The spline interpolation expression is: Hermite basis functions .in, For node function values, Let be the nodal derivative (i.e., the slope at the interval endpoints). Monotonicity constraints are expressed using linear inequalities: for any node, must satisfy... and To avoid oscillations caused by excessively steep derivatives, additional constraints are applied. ,constant Take between 2 and 5. I won't go into details.

[0033] The node-solving problem is formulated as a constrained quadratic programming problem, with decision variables being vectors. and The objective function is ,in This represents the original set of quantiles used for constraints (e.g., 25%, 50%, 75%, 90%). Weights corresponding to the reference quantile values Based on the sample size or confidence level of this quantile in the anchor set, the smoothing regularization term penalizes derivative oscillations, and the prior term... Provides conservative backoff (prior can be an identity mapping or parameters from the previous batch). The constraint set includes monotonicity and upper bound of derivatives, and adjustable range constraints for node positions. (When fine-tuning of node positions is allowed), and numerical stability constraints. .

[0034] When the anchor sample size is finite, the contraction strategy is activated: if the number of samples in a certain quantile group is... Then the weight according to Reduce, and shift the target value of the node towards the prior. To win over, or to smooth out the regularization coefficient Increase to the preset upper limit. Solving the listed constraint quadratic programming problem can be performed in parallel on GPU / CPU using the standard QP solver. If infeasibility is found during the solution process, the fallback strategy is to perform isentropic monotonic regression on Y to obtain a monotonic least squares solution, and then use finite difference estimation to estimate the initial value of M.

[0035] After obtaining the node parameters, perform a full analysis of all the original scores. Calculate calibration score Define the mapping residual metric. Perform iterative calibration: if And the number of iterations is less than Then use the current Recalculation of anchor points Update target The weighted terms are used to solve the new QP; when or reach The process terminates at the designated time, and the parameters are recorded in the batch file.

[0036] Fixed judgment threshold In the calibration domain, the quantile constraint is determined. A common strategy is to take the first reference anchor from the set of reference anchors. Percentile calibration score. Parameter Based on an acceptable recall / precision tradeoff (e.g., 95%), when the anchor sample size is insufficient to reach the betting threshold, a prior threshold is linearly mixed with the current batch threshold, with the mixing coefficient proportional to the anchor sample size.

[0037] In implementation, to avoid scoring bias caused by traditional algorithms relying solely on visual differences, this step constructs a monotonically piecewise mapping function based on quantile constraints, using physicochemical content as a reference. Samples with a content ≥0.058% of verrucoside glucoside are designated as the high-confidence interval, corresponding to the upper segment of the mapping function; samples in the 0.027%–0.058% range are designated as the medium-confidence interval; and samples below 0.027% are designated as the low-confidence interval. The system introduces a chemical content gradient term into this function, forming a continuously differentiable score correction curve. When the visual characteristics of a sample are inconsistent with its chemical indicators, the mapping function automatically adjusts the bias Δθ, ensuring monotonically consistent scoring results. This method ensures that although different batches and sources of medicinal materials may have visual differences, a unified judgment standard can still be obtained after correction based on the chemical benchmark.

[0038] To verify the effectiveness of the calibration method, the classification accuracy was compared between two approaches: traditional fractional mapping and the introduction of a physicochemical calibration model. The results showed that the overall accuracy without physicochemical calibration was 87.3%, while the accuracy increased to 96.1% using the method of this invention. Particularly in samples with uneven illumination and significant color deviations, the misclassification rate decreased from 12.7% to 3.9%. This demonstrates that the introduction of physicochemical data significantly improves the stability and robustness of the model in complex sample environments.

[0039] Step 6: Calculate the structural tensor direction and the difference in cross-seam strength on both sides of the seam strip to generate the seam continuity field and establish a boundary mask. Specifically, this step receives the seam strip centerline L and sub-window index generated in step 1, and performs discrete sampling of the seam strip along the short axis direction. It calculates the continuity metric of each sampling point and generates a spatialized mask for subsequent reliability ranking of the cross-window adjacency graph and seam neighborhood removal.

[0040] The working principle is as follows: a sequence of uniformly discrete sampling points is performed on the centerline L along the seam strip. The sampling step size is recorded as (pixels), the recommended value range is 1 / 20 to 1 / 5 of the sensor resolution. For each sampling point, local sub-regions are cropped to the left and right along the short axis. Sub-region size set to (short axis width) For a multiple of several sampling step sizes, the major axis (Covering several adjacent sampling points), the sub-regions are rotated and aligned relative to the center line in the same direction to ensure that texture statistics are compared in the same local coordinate system.

[0041] Local sub-region grayscale image First calculate the horizontal gradient with vertical gradient The formula is: To suppress noise, the squared gradient term is Gaussian smoothed to construct the structure tensor field. Among them, Gaussian kernel ,parameter Recommendation Pixel range, increase It brings stronger smoothness and scale invariance but sacrifices local orientation sensitivity.

[0042] For each pixel, calculate the eigenvalues ​​and principal direction of the 2×2 symmetric matrix J. The eigenvalue expression is as follows: main direction use ,in Return angle at Interval; due to the 180° period of the direction, the subsequent spatial average is represented by a double angle (the following average of the direction uses the double angle method). Construct the local consistency coefficient. The coefficient ranges from [0,1). A value close to 1 indicates strong local anisotropy of texture (with a clear main direction), while a value close to 0 indicates isotropic texture or noise.

[0043] For each sampling point, the gradient intensity field is statistically analyzed in the left and right sub-regions. The mean and robust standard deviation are defined as follows: Similarly, we can conclude that The normalized cross-joint strength difference is defined as... This quantity measures the degree of deviation in strength distribution on both sides. The larger the value, the more severe the sudden change in strength across the joint, which is commonly seen in joint artifacts or material fracture.

[0044] The directional difference is calculated by directional weighting within the left and right sub-regions, for each pixel direction. Calculate double angle vectors According to its corresponding consistency coefficient Weighted summation restores the average direction: , respectively obtained . direction difference Standardization to Post-construction direction consistency factor The factor is set to 1 to indicate that the directions are consistent, and 0 indicates that the structure is orthogonal or strongly inconsistent in the opposite direction.

[0045] The smaller of the left-right consistency coefficients is taken and combined with the direction factor and intensity attenuation to form the sampling point continuity coefficient. If both sides have stable main directions and consistent directions with small intensity differences, a high CF value indicates good continuity at the joint. Conversely, a low CF value indicates that the joint may be an artifact or an actual fracture, and the reliability weight of this joint segment needs to be reduced in subsequent processing.

[0046] Integral along the centerline to accumulate continuity lines This scalar reflects the total continuity of the entire joint segment; the smaller the value, the less reliable the entire joint segment is. Threshold setting rule: If or more than the seam sampling points This triggers the unreliable joint flag. Typical parameters: This is the empirical value of the 10th percentile within the sample. Take a value of 0.2–0.5. Use a value of 0.2–0.4, which can be adjusted according to the material texture and on-site noise.

[0047] The process for generating a binary boundary mask for the continuity coefficient field of the sampling points is as follows: First, thresholding is performed on the CF map to extract the low-value range. Marked as discontinuous segments, then morphological closing operations are performed to fill small gaps, followed by removing segments with areas smaller than [the specified area]. The isolated fragments were then subjected to a dilation process on the mask to obtain the width. The seam influence zone. The recommended size for the morphological structural element is 3–11 pixels, depending on the width of the sub-window and the seam. The mask output fields include the center position, length, local minimum CF value, and percentage of each discontinuous segment. These are used in step 7 to perform reliability sorting and weighted suppression of adjacent edges across windows, and in step 2 to accurately remove seam neighborhoods when filtering anchor areas.

[0048] As one possible implementation method: gradient and Gaussian convolution are accelerated using separable filtering and integer convolution; pixel features of the structure tensor are stored using single-precision floating-point and... Upper / lower bound pruning is performed to avoid numerical overflow; directional averages are represented using double-angle notation to address 180° periodicity; statistics Incremental windows or integral plots are used for rapid updates, facilitating online processing. If the seam covers a large portion of the image area, or... If the value remains low for multiple consecutive frames, a higher-level domain drift process is triggered (see the domain drift determination in step 5), such as reconstructing the seam index or adjusting the threshold set. The specific strategy can be set according to the process requirements. It should be noted that the structural tensor consistency coefficient in this application characterizes the directional clarity of the local texture and provides directional average weighting; the directional consistency factor can uniformly measure the reverse and forward directions.

[0049] Step 7: Within each sub-window, perform connected component labeling based on the calibration score and threshold, and map it to global coordinates. Construct a cross-window adjacency graph composed of seam continuity and shape compatibility, and generate a global defect instance set through merging and nonmaximum suppression. The specific working principle is as follows: It starts from the sub-window level. For any sub-window i, the input is the calibration score field corresponding to the sub-window. (Outputs from steps 4 and 5) and the global decision threshold (Step 5 definition). Sub-window binarization uses the following rule: binary mask. To suppress isolated noise, after binarization... Perform morphological opening and closing operations (structural element radius) (A resolution of 1–5 pixels is recommended, but can be adjusted based on resolution). Connectivity component labeling uses the 8-connectivity rule to obtain a set of local candidate regions. For each connected component Computational geometry and statistical properties: Area ,perimeter Boundary polygon (outline) Center of gravity convex hull area entity rate The spindle length ratio AR (based on the second-order central moment), Hu invariant moment, etc., are calculated simultaneously, along with the average calibration score within the region. .like or or If the defect is removed, the remaining area is considered as a candidate for local defect.

[0050] The local to global coordinate mapping uses the sub-window indexing and registration transformation from step 1. Let the pixel coordinates of the origin (top left corner) of sub-window i in the original frame be... If the sub-window undergoes affine transformation during registration... (For example, the affine matrix generated by sub-pixel registration), then the boundary points of the connected components will be... Mapped to global coordinates The centroids, boundaries, and polygons of connected components are all transformed using the above mapping to obtain the global initial node set. Each node Vector with attributes .

[0051] Window adjacency graph The construction is based on spatial proximity, overlap / proximity, seam continuity, and shape compatibility. For any node pair... (Usually limited to adjacent child windows or global distance less than) Calculate the following metrics for the node pairs and normalize them to [0,1]: Shape compatibility It uses the Intersection over Union (IoU) ratio as its core and supplements it with boundary similarity. Definition , , (where the length is the diagonal of the boundary), then take ,in A positive weight is used to control the penalty for differences in boundary and shape proportions. This design balances area overlap with contour consistency.

[0052] Seam continuity factor Take the mean of the continuity function CF (output of step 6) along the ridge segment crossing the seam between the two nodes. If the segment belongs to the "unreliable zone" of the seam mask, then let Reduce weights (multiply by a factor of 0–1). Set after normalization. .

[0053] Score similarity Alternatively, a correlation coefficient can be used to reflect the consistency of calibration scores within the region.

[0054] The above sub-items are linearly or convexly combined to form edge weights. parameter and normalize it. Typical value range: ,and When any node crosses the seam and the seam mask is marked as strongly discontinuous, it is forced to... Reduce the value to a low value or set it to zero to avoid accidental merging across seams.

[0055] The merging strategy uses a "restricted disjoint-set data structure" to merge edges in descending order of weight on the weighted graph: all edges are merged according to... Sort the edge set in descending order, traverse the edge set, and assign weights. The edges attempt to merge the sets to which the nodes at both ends belong; constraints are checked before merging: area of ​​the merged set. Not exceeding (To prevent mistakenly merging large backgrounds as defects), merging must not cross unreliable segments with seams (i.e., if the merging path must cross seams marked as "unreliable," then merging is rejected), and the average shape compatibility of the merged set must not be lower than [a certain value]. If the constraints are met, perform a union operation and update the set properties (area, boundary, average score, convex hull, etc.). This union-find set strategy is equivalent to merging based on local reliability, which can retain high-confidence cross-window instances while avoiding erroneous merging caused by seam artifacts.

[0056] After unifying the search set and obtaining the initial global candidate set, weighted intersection-union ratio (IURR) nonmaximum suppression is performed. The weighted IURR of two candidate instances A and B is defined. ,in The average CF corresponding to the overlapping region of the two instances. The adjustment coefficient is used. The NMS process is as follows: Candidates are sorted in descending order of their confidence scores (e.g., the ensemble average calibration score plus area-weighted sum). The candidate set is traversed, and for the current candidate c, its maximum value with the set of selected instances is calculated. ;like Then compare scores: if score c > the maximum score of the selected instance that intersects with it, then keep c and remove all instances that intersect with it. Instances of c; otherwise, discard c. This rule is equivalent to "retaining the highest scorer in the local overlap group, while allowing newcomers to replace previously selected ones when necessary," thus balancing detection sensitivity and deduplication effectiveness. Parameters The initial engineering recommendations are 0.5–0.7, 0.3–0.6, 0.5, and 0.5, which can be adjusted according to materials and false alarm tolerance.

[0057] When the number of initial candidates is abnormally large Or the total number of instances after merging < At that time, the activation threshold is adaptive: increase Increase Or switch to a more conservative approach. Parameters; if the seam continuity field is below the threshold on a large number of adjacent edges, backtrack to step 6 to reconstruct the seam mask or trigger the domain drift notification in step 5.

[0058] The final output is a global set of defect instances. Each instance contains: a global polygon mask, boundary chains, area, average calibration score, shape features, cross-window component node indexes, and a seam reliability summary. This output serves as input for steps 8 / 9 (attribution mapping, carrier-level statistics) and is also written to the batch archive for retrospective fine-tuning of mapping parameters for subsequent batches.

[0059] Step 8: Generate the foreground response map of the medicinal herb slices. Obtain the set of slice instances using distance transformation and watershed methods, and repair the seam contour according to the seam continuity field cost; please refer to [link / reference]. Figure 3 The diagram shown illustrates foreground response and watershed correction. The input for this step is the original registration frame image. Step 7 outputs a global defect instance mask set. The seam continuity field CF(x) output from step 6 (defined on the seam ridge or its neighborhood). Foreground enhancement first removes static background noise using the temporal background model B(x,y): the median of the pixels in the nearest K frames is used as the background reference. Or use a weighted recursive background As an alternative; parameter K is recommended to be 3–7. A value of 0.85–0.98 is recommended, but this can be adjusted based on transmission rate and scenario variations. The original difference in foreground response is defined as... Then, local normalization filtering and contrast enhancement are performed: local mean is calculated. Standard deviation (window radius) (typically 8–32 pixels), to obtain the normalized response And apply restricted histogram equalization (CLAHE) or Gaussian smoothing (radius) to R. To generate the final foreground response map To suppress the influence of the global defect mask, first... Regions are masked during background modeling or weighted less in the final foreground response to avoid defective areas interfering with sheet segmentation.

[0060] based on Generate binary foreground mask The steps are as follows: Use Otsu or adaptive local thresholding get And on Perform morphological cleanup (opening operation for noise reduction, closing operation for hole filling, structuring element radius adjustment) The distance transformation uses Euclidean distance transformation (EDT) to calculate the distance from each foreground pixel to the nearest background pixel. This transformation can be implemented using Felzenszwalb's linear-time EDT algorithm or the classic Euclidean DT. Interior point seed set. The set of local maxima from D: First smooth D (Gaussian radius) ), then extract the satisfaction {if and only if} .

[0061] in Set the minimum interior point distance threshold (it is recommended to be proportional to the minimum radius of the target, such as 4–12 pixels). For local neighborhood windows (radius) At the same time, nonmaximum suppression ensures that the distance between any two sub-subs is greater than 1. (Avoid overly fine segmentation). Apply a rejection rule to seeds near the seam mask area: if the seed is far from the seam centerline... And corresponding seam points If the seed is not found, it will be removed to prevent seam artifacts from generating incorrect slices.

[0062] Watershed segmentation uses a gradient map G(x,y) as a damping field. The gradient is calculated from the input brightness or foreground response and is defined as follows: Or use nonlinear transformation of edge magnitudes Adjust sensitivity to deep textures. Mark watersheds to... To label the fragments, run the label watershed algorithm to obtain a preliminary set of fragment instances. For each Calculate area ,boundary Average prospect response Remove or The pseudo-film.

[0063] When the watershed boundary crosses the seam neighborhood (i.e., the boundary overlaps with the seam influence zone), calculate the cost of that boundary segment. in For discrete points of the boundary segment, The sampling step size is used to construct the joint projection continuity field. Diffusion of CF along the ridge line into a two-dimensional neighborhood: in Control the impact bandwidth and define pixel cost mapping The boundary replacement problem is transformed into a problem involving a graph composed of pixels. Find the minimum cost path. For the endpoints p and q of the watershed boundary to be repaired (located at both ends of the boundary segment), solve using Dijkstra's algorithm or A* on the subgraph containing the joint influence zone. If the cost of the replaced path satisfy To replace the penalty term and avoid minor oscillations, it is recommended to use a term related to boundary smoothing. Then, replace the original boundary segment with p*, followed by boundary smoothing (B-spline or curve fitting) and pixel-level redistribution. Merge or segment adjacent sheets to restore a coherent sheet mask. This repair process can be performed in parallel on all cross-seam boundaries, or preferentially on local segments where low seam continuity is found. The iteration criterion is: if the replacement leads to a decrease in the total boundary cost... Then continue iterating, or until the maximum number of iterations is reached. Stop; Recommend .

[0064] After boundary repair, merge adjacent overlapping sheets. And apply hole filling and small piece removal (area threshold) Output the final collection of slice instances. The modified boundary mask set, with each piece having a repair summary field (whether it is a cross-seam repair, the original cost, the new cost, and the cost reduction amount), is written to the batch file for auditing and subsequent parameter fine-tuning.

[0065] It should be noted that the foreground response map represents a normalized and enhanced continuous response field, which is beneficial for stabilizing seed detection and gradient calculation; the interior point seed refers to the local maximum of the distance transformation, which serves as a watershed marker to ensure that the segmentation granularity is consistent with the physical sheet; the seam continuity field (CF) is different from a simple seam mask. It is a field with a numerical continuity metric, which can be projected onto the two-dimensional pixel domain and drive the minimum cost path algorithm.

[0066] Step 9: Construct a bipartite graph of defect instances and sheet instances. The edge weights are given by the region intersection-union ratio, boundary direction difference, and seam continuity integral. Solve for the maximum weight matching to obtain the defect attribution mapping. Specifically, in the candidate pairing generation stage, for any defect node... With sheet body node First, calculate the Euclidean distance between the centroids. Only for those that satisfy The pairing constructs a candidate set (where (Can be set according to typical film size or pixel scale). Three types of raw quantities are calculated for each candidate pair: The first type, the intersection-union ratio of regions, is defined as... in This represents pixel area or a continuous area measure. To compensate for the difference in expected overlap caused by area imbalance, a kernel regression based on batch anchor samples is used to estimate the expected overlap function. Let the area ratio... Batch anchor sample set Estimate using Nadaraya–Watson kernel regression: The kernel function is a Gaussian kernel. The bandwidth h can be determined according to the Silverman rule or cross-validation (in engineering practice, it can be initially set to 0.05–0.2). The normalized adjusted intersection-union ratio is denoted as... in To prevent zero items.

[0067] The second type calculates boundary normal similarity using a weighted cosine average of the normal vectors at contact / coincident boundaries. This involves sampling a set of points at the overlapping or nearest-neighbor boundaries of the two regions. Calculate the boundary normal vector of the defect and the sheet body at each point. (The contour tangent is estimated using the difference between adjacent points, and the normal is the tangent rotated counterclockwise) (and normalize), and take the joint continuity coefficient output from step 6. As a weight, we define weighted cosine similarity. The value of this quantity is approximately in the range of [-1, 1]. For subsequent linear combinations, it needs to be mapped to [0, 1]. This can be achieved using... .

[0068] The third type involves sampling the line integral of the seam continuity along the coincident boundary or crossing path between the two. in The sampling step size (pixels) is numerically normalized to average continuity, reflecting the reliability of the joint across the boundary segment.

[0069] After obtaining the three types of raw quantities, construct the training matrix. (M is the number of high-confidence anchor samples within the batch), each row contains The normalized value of the target vector Indicate the true / false labels for the pairings (true matches with high confidence are automatically labeled as 1, or manual / historical matches are used as supervision). Solve the restricted least squares problem, i.e.: This problem can be solved using nonnegative least squares or quadratic programming with equality constraints (numerical solution: active set method or using the standard QP solver); if the sample size is insufficient, NNLS can be used. Post-normalization The resulting weight vector It is the linear combination coefficient of the three types of quantities, and is archived with each batch to achieve batch adaptation.

[0070] The edge weights are constructed as a linear combination: To avoid the creation of isolated, low-quality edges, a dynamic thresholding rule is adopted: the edge weight distribution for a certain defect or local candidate set. Calculate the median With absolute median Take dynamic threshold where constant Version 1.0–2.0 is recommended. This is a preset lower limit (e.g., 0.1). Only when... and When constructing edges on a bipartite graph and empower .

[0071] Constructed bipartite graph The maximum weighted matching algorithm is used to find the optimal one-to-one assignment. The graph is then transformed into a weight matrix. For entries without edges, set the weight to 0 or a minimum value, and use the Hungarian algorithm or a general maximum weight matching algorithm to solve the problem of maximizing the total matching weight. To handle cases with unequal numbers of nodes, add virtual nodes with weights of 0 on the side with the smaller weight, thus obtaining one-to-one matching or partial matching results.

[0072] Detailed rules for matching: If a defect Two optimal candidates and suboptimal Matching weight difference Less than the discrimination threshold This is considered a fuzzy match, and the integrals of the seam continuity lines are compared sequentially according to priority. Boundary normal similarity Make a final attribution decision; if it still cannot be determined, mark the pairing as "difficult," write it to the batch file, and trigger manual review or further judgment based on subsequent frame accumulation. Default Take 0.05–0.15.

[0073] Once all affiliations are determined, a mapping set is formed. And calculate the attribution confidence (e.g.) (Ranking after normalization), unmatched defects are retained as isolated instances and marked as "unassigned"; if a piece is matched by multiple defects, a one-to-one match ensures uniqueness. If it is necessary to support scenarios where a piece corresponds to multiple defects, multiple edges with lower thresholds can be added to the same piece in the post-processing stage and multiple assignments can be recorded.

[0074] Step 10: Generate defect statistical labels based on the defect attribution mapping, write the batch domain descriptor, mapping parameters, reference quantile group, decision threshold, and seam continuity summary into the batch file, and monitor drift trigger parameter updates within a sliding time window. In implementation, defect category counting... Summarized by category index. Defect area percentage. The percentage of this type of defect area to the total foreground area of ​​the medicinal material: The centroid spatial distribution is based on the sample mean vector. With covariance matrix The principal direction and eigenvalues ​​of the covariance are represented as label fields. The slice overlap rate is defined as the defect coverage rate at the slice level: for each slice... Calculate coverage ratio The overall overlap rate is archived in quantile statistics (25%, 50%, 75%) and mean form. Defect continuity scoring uses normalized online joint continuity integration: for each defect instance... Calculate the set of boundary sampling points Continuous scoring The minimum, mean, and standard deviation are recorded simultaneously as quantitative indicators of subsequent segmentation and repair quality. All of the above statistics are used to form a vector tag set in the batch-level summary and written into the archive.

[0075] Batch archives are organized by timestamp and version number, and each record contains a set of fields: batch number, timestamp, and mapping parameter snapshot. Reference quantile information, judgment threshold Batch domain descriptor The archive includes a seam continuity summary (e.g., seam line integral distribution), anchor sample statistics, trigger flags, and manual verification annotations. The archive uses a verifiable format (JSON or relational table rows) and calculates the signature hash and version chain number upon writing to support rollback. Specific details can be determined based on actual needs.

[0076] Drift monitoring is performed on a sliding window frame, where the sliding window contains either the most recent W batches or the most recent N tens of thousands of frames, and the monitoring sequence includes a set of key quantiles in the anchor region. (e.g., 25%, 50%, 75%, 90%) and empirical distribution of calibration scores. The test statistic is defined as follows: Significance level Recommended value: 0.01 or 0.05. The distribution difference is then determined to be significant, triggering an update. CUSUM change point detection is used to detect mean shift: [The text then abruptly shifts to a different topic:] ...take the sequence... Let the target mean be the anchor median or a certain quantile. Buffer constant k, decision threshold h, rising cumulative cumulative decline ,when or Alarm. Formula for tracking Exponentially Weighted Moving Average. Control limits adopt The corresponding lower bound and the recommended parameter range The monitoring logic is as follows: any statistical test exceeding its confidence threshold triggers a parameter update process.

[0077] The parameter update process is based on constrained weighted incremental least squares fitting. Abstractly, it involves the mapping function parameter vector... Solving the weighted least squares problem: objective function Where X is the anchor sample feature matrix (the original score of the anchor, etc.), y is the target mapping value (referencing the quantile target), and the weight matrix... Reflecting sample confidence and time decay, the decay strategy employs exponential decay. Attenuation factor Recommended value: 0.90–0.995; Ridge regularity coefficient Choose from 1e-4 to 1e-1, adjusting according to the number of problem conditions. For constraints, monotonicity and nodal slope constraints can be written as linear inequalities. The solution process employs incremental updates: if a new batch of observations provides matrix increments... Solve using regularized normal equations Incremental updates of the inverse matrix are performed using the Sherman–Morrison–Woodbury formula to improve online performance; when constraints are active, a constrained quadratic programming solver (active set method or projected gradient method) is used. After solving, the mapping function is simulated and replayed: the new parameters are applied to historical anchor samples to calculate the new calibration score sequence. Thresholds are determined based on quantile reestimation. And introduce smoothing to the threshold. .

[0078] After the update, key performance indicators (recall, precision, false positive rate) are calculated on a short-term validation set or retention window. If any key indicator decreases by more than a preset threshold compared to before the update, it is marked as a failure and automatically rolled back to the previous parameter. Rollback records include trigger time, trigger statistic, sample summary, version number, and rollback flag. All change records are written to the batch file with the following fields: old parameter snapshot, new parameter snapshot, trigger reason, test statistic, sample summary, version number, rollback flag, and manual review annotation.

[0079] In the design of this invention, the aforementioned physicochemical standards and grading information are written into the batch file for long-term drift monitoring and automatic updates. Specifically, the system statistically analyzes the average content of isoflavone glucoside and total flavonoids in 30 consecutive batches within a sliding time window. When a trend deviating from the predetermined reference range (e.g., for top-grade samples, d ≥ 1.4 cm, total flavonoid content ≥ 6.2954%) is detected, the model retraining and threshold resetting process is automatically triggered. In this way, the system can adaptively track changes in physicochemical properties caused by the source of the medicinal material, cultivation environment, or storage conditions, maintaining consistency between the detection threshold and the actual quality of the medicinal material.

[0080] In the batch domain descriptor storage, the system also stores the corresponding average physicochemical content. and variance And based on this, a drift statistics matrix is ​​constructed. When the Euclidean distance between M(t) and the historical mean M0 Exceeding the preset threshold At this time, the system automatically executes a self-closed-loop correction process of "fractional domain remapping - threshold recalibration - descriptor recalibration". This mechanism effectively prevents recognition bias caused by model aging and ensures the stability and traceability of long-term monitoring of defects in Chinese medicinal materials. It should be noted that the batch archives in this application can fully record batch-level domain descriptions, mapping parameters, thresholds, seam summaries, trigger logs, and version history for reference in traceability, auditing, and subsequent fine-tuning. Drift triggering refers to the parameter update triggering process determined by a combination of sliding window statistical tests.

[0081] Finally, one possible implementation of the image acquisition method for consecutive frames and overlapping windows in this invention is as follows: camera resolution ≥ 2048×2048, sub-window size preferably 128×128 pixels, sub-window overlap rate r is 0.10–0.30 (0.20 recommended); frame rate and transmission line speed must meet pixel displacement constraints. ,exposure According to motion blur limit Pixel calculation; the seam width is set according to the ratio of α to the short side of the sub-window, with α ranging from 0.05 to 0.15 (0.10 recommended). The purpose of this imaging method is to ensure that films of different thicknesses and colors appear at least once in several consecutive frames and overlapping sub-windows, facilitating subsequent anchor point association across windows and frames.

[0082] As one possible implementation method, in the specific implementation of the machine vision-based method for detecting defects in traditional Chinese medicine materials described in this invention, Ligusticum chuanxiong was selected as a representative medicinal material for joint verification experiments. Ligusticum chuanxiong has a typical rhizome structure, and its appearance, internal components, and storage stability vary significantly, which can comprehensively reflect the adaptability of the method in morphological discrimination, optical feature recognition, and batch drift calibration.

[0083] A total of 21 batches of Ligusticum chuanxiong samples from different origins and storage times were collected for the experiment. The content of ligustilide I and the transfer rate of ligustilide in the extract were determined according to different origins. The samples were divided into a fresh sampling group and a stored group based on storage time; the fresh sampling group consisted of newly harvested samples, while the stored group covered samples stored for 1-8 months. To ensure image acquisition consistency, a high-resolution industrial camera was used for continuous frame shooting at a resolution of 4096×2160 pixels. The light source was a constant color temperature 5000K area array light source, and the distance was fixed at 45cm. Images were segmented using a sliding window with a window size of 512×512 pixels and an overlap width of 20%. A seam index was established for subsequent continuity correction.

[0084] Place of origin Content of ligustilide I in raw medicinal material Transfer rate of Ligusticum striatum extract I Common appearance types B01 0.9 70% Nodular, fist-shaped clumps, 5–6 cm in diameter. The surface is brown, rough, and wrinkled, with a sunken, nearly circular stem scar at the apex. The cut surface is yellowish-white and nearly circular, with distinct wavy rings and dense yellowish-brown oil. B05 1.1 74% It is multinodular and spindle-shaped, with numerous parallel raised nodes at the top. The surface is grayish-brown, rough, and wrinkled, with a sunken, nearly circular stem scar at the apex. The cut surface is yellowish-white and irregularly shaped. It has distinct irregular textures and scattered yellowish-brown oil dots, fewer than those in B01 and B05. B09 0.22 52% Nodular, fist-shaped clumps, 5–6 cm in diameter. The surface is brown, rough, and wrinkled, with a sunken, nearly circular stem scar at the apex. The cut surface is yellowish-white and nearly circular, with distinct wavy rings and dense yellowish-brown oil. B13 0.34 56% It is multinodular and spindle-shaped, with numerous parallel raised nodes at the top. The surface is grayish-brown, rough, and wrinkled, with a sunken, nearly circular stem scar at the apex. The cut surface is yellowish-white and irregularly shaped. It has distinct irregular textures and scattered yellowish-brown oil dots, fewer than those in B01 and B05. The content of ligustrolactone I in each batch of medicinal slices was determined, and time-series data were generated. The measured data are shown in the table below: Experimental results showed that the content of ligustilide I in Ligusticum chuanxiong from different origins and the transfer rate of ligustilide I in the extract varied significantly. The content of ligustilide I changed over different storage periods. Both the content and the transfer rate of ligustilide I in the extract showed an increasing trend in the first six months, with a faster increase in content between 4 and 6 months. After 6 months of storage, both levels tended to stabilize.

[0085] Based on the above experimental data, to further verify the effectiveness of the machine vision-based method for detecting defects in traditional Chinese medicine in terms of physicochemical index calibration, the change in the content of ligustrazine I was used as the main external reference parameter to construct a batch-domain reference curve. Experiments showed that the content of ligustrazine I varied within the range of 0.22–1.10 mg / g in ligustrazine samples from different origins and with different storage periods. The content increased significantly in the first six months of storage and then stabilized after six months. This trend was used to establish a quantile mapping model to correct the bias caused by storage effects on visual scores. Through quantile modeling, the ligustrazine I content at 0.32 mg / g, 0.46 mg / g, 0.61 mg / g, and 0.69 mg / g was mapped to the 25%, 50%, 75%, and 90% nodes of the score domain, respectively, forming a batch-domain calibration baseline. This mapping was used to correct score fluctuations caused by storage time or surface color differences, ensuring that the score domain distribution reflects the true changing trend of the intrinsic active components of the medicinal material.

[0086] In the image preprocessing stage, the system performs sliding window segmentation on consecutive frames of all captured samples and filters based on gradient variance, texture energy, and color entropy to remove splicing gaps and boundary interference regions. To ensure the stability of feature extraction, regions with rich details and uniform reflection features are selected as anchor regions. The selection criteria are shown in the table below: index Threshold (quantile ratio) illustrate gradient variance > 70% percentile Preserve rich details in the area Texture Energy > 60% percentile Preserve areas rich in structural information Color Entropy < 90th percentile Excluding high-noise chroma regions The selected anchor regions were used to extract color histogram statistics, power spectral moments, and gray-level co-occurrence moments, and combined with storage period information to generate batch domain descriptors. The descriptor has a dimension of 128, incorporating time-series variation parameters corresponding to the content of ligustilide I to ensure comparability of feature differences between different batches after standardization.

[0087] All sub-windows are processed by the defect detection model to generate an original score set. To eliminate the systematic offset of scores due to storage conditions, this invention introduces a monotonic piecewise mapping function based on batch domain descriptor differences to map the original score set to the calibration score space. The mapping nodes are shown in the table below: quantiles raw fractions Corresponding content of Ligusticum lactone I (mg / g) calibrated scores 10% 0.12 0.30 0.11 25% 0.20 0.32 0.17 50% 0.35 0.46 0.30 75% 0.58 0.61 0.46 90% 0.81 0.69 0.68 This piecewise mapping function automatically adjusts the weights during the intermediate storage phase when content changes significantly, ensuring that samples with high content correspond to higher score responses, while samples whose surfaces become dark or oil spots become blurred due to storage still maintain the correct ranking relationship after calibration. After calibration, the average score of samples after two months increased by approximately 7% compared to the initial samples, samples after four months by approximately 18%, and samples after six months by approximately 20%, demonstrating that the model effectively eliminates false alarms caused by changes in appearance.

[0088] During image stitching, the seam area is a concentrated area of ​​stitching errors across multiple frames, often leading to structural pseudo-cracks. Therefore, this embodiment calculates the principal vector difference of the structural tensor across the seam. and strength difference Define continuity The system generates a continuity field based on this. Regions with a continuity value below 0.5 are automatically enlarged to a mask half-width of 20px to avoid false detections. This continuity field is also used in subsequent segmentation and repair processes. When the watershed algorithm encounters a fracture boundary, the system automatically repairs it based on the shortest path constraint of the continuity field, enabling the differentiation between crack-like defects and natural texture fractures.

[0089] During the matching phase between the sheet body and the defect instance, the system constructs a bipartite graph for the sheet body and the defect instance. The edge weight w is determined by the region intersection-union ratio (IoU), the boundary direction difference Δφ, and the seam continuity integral. Composed of: .

[0090] Parameters α=0.5, β=0.3, γ=0.2.

[0091] To adapt to the typical morphology of Ligusticum chuanxiong, a "shape compatibility S" constraint term is added. Matching is rejected when S is below 0.6. The shape compatibility is derived from a typical fist-shaped sheet contour template, with shape frequency domain features extracted using Fourier descriptors. This template originates from the morphological measurement results of the medicinal material, ensuring that the attribution logic conforms to the natural morphological distribution pattern of Ligusticum chuanxiong.

[0092] To verify the stability of the calibration strategy, the average calibration score and the change in the content of ligustrazine I in samples with different storage periods were compared and analyzed. The results are shown in the table below: Storage time (months) Average calibration score Content of Ligusticum lactone I (mg / g) Fractional drift Δ 0 0.18 0.30 0 2 0.25 0.46 +0.07 4 0.36 0.61 +0.18 6 0.49 0.68 +0.31 8 0.62 0.69 +0.44 The results showed a significant positive correlation between fractional drift and changes in the content of ligustrolactone I. The system was designed with drift triggering conditions within a sliding time window: when the median calibration score deviated from the baseline by more than 12%, while the change in ligustrolactone I content was less than 0.01 mg / g, model drift was identified, and automatic re-estimation of the mapping function was triggered; otherwise, it was considered a natural physicochemical change in the medicinal material. This strategy enabled the system to distinguish between storage-induced appearance degradation and model drift, ensuring long-term robustness.

[0093] To evaluate the consistency of the model output, 100 sets of samples were selected, and the defect areas were determined by human experts and compared with the model output. The consistency rate is as follows: sample group Number of defects Consistency Decision Number Consistency rate (%) G1 12 11 91.7 G2 15 14 93.3 G3 18 17 94.4 G4 14 13 92.9 average — — 93.1 Comparative results show that the model's detection results are highly consistent with manual judgment standards under different storage periods and appearance conditions. Through the synergistic effects of feature extraction, calibration mapping, and shape restoration, the system can accurately reflect the visual response patterns of *Ligusticum chuanxiong* during physicochemical changes. Experimental results regarding storage period and *Ligusticum chuanxiong* lactone I content provide a reliable basis for fractional calibration and drift monitoring, while the combination of morphological templates and texture descriptors ensures the model's adaptability and stability in structurally complex samples. Overall, the results verify the robustness and consistency of the method under different batches and physicochemical conditions, demonstrating its promising industrial application potential in the field of intelligent quality detection of traditional Chinese medicine.

[0094] It should be noted that the mathematical formulas, derivations, symbol definitions, and parameter calculation methods used in this specification are all for the purpose of further clarifying and demonstrating the technical content of this invention, so that those skilled in the art can more intuitively and accurately understand the working mechanism and technical effects of this invention. These formulas are only used as quantitative expressions or illustrative examples of technical features and do not constitute limiting conditions of the claims of this invention. Those skilled in the art should understand that, without changing the core idea of ​​this invention, the parameter forms, calculation methods, numerical ranges, and even symbol representations involved in the formulas can be equivalently replaced or simplified in engineering according to the actual application environment. The specifics can be determined according to the actual situation, and no limitation is imposed. It should also be emphasized that the formulas in this specification are not theoretical derivations in the style of academic research papers, but rather an engineering description of the embodiments of this invention. Their purpose is to enhance the understandability and implementability of this invention, rather than to increase redundancy and complexity. Those skilled in the art can choose whether to use such quantitative tools when reading this specification, or can achieve the same technical effects through other equivalent methods.

[0095] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these specific embodiments are merely illustrative. Those skilled in the art can omit, substitute, and modify the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, combining the above method steps to perform substantially the same function and achieve substantially the same result according to substantially the same method falls within the scope of the present invention. Therefore, the scope of the present invention is defined only by the appended claims.

Claims

1. A machine vision-based Chinese medicinal material defect detection method; characterized in that: The method comprises the following steps: Step 1, collect continuous frame images of Chinese medicinal materials, generate a set of medicinal material sub-windows according to a fixed window and an overlapping width, and establish an index of medicinal material joint seams; Step 2, in the set of medicinal material sub-windows, screen a set of Chinese medicinal material anchor regions according to gradient variance, texture energy and color entropy, and remove medicinal material joint seam neighborhoods; Step 3, extract Chinese medicinal material color histogram statistics, power spectrum moments and gray level co-occurrence matrix of the set of Chinese medicinal material anchor regions, construct a Chinese medicinal material batch domain descriptor, read a reference Chinese medicinal material descriptor and a reference quantile group, and perform Chinese medicinal material defect scoring on all medicinal material sub-windows to obtain an original score set, and count reference region quantile points as a calibration reference; Step 5, based on the difference between the Chinese medicinal material batch domain descriptor and the reference Chinese medicinal material descriptor, construct a monotonous segmented mapping of the score domain, set quantile constraints, obtain mapping parameters, calibrate the original score, and set a fixed judgment threshold in the calibration domain; Step 6, calculate the structure tensor direction and the cross-seam intensity difference on both sides of the medicinal material joint seam to generate a medicinal material joint seam continuity field and establish a boundary mask; Step 7, based on the calibrated score and the threshold, perform connected component labeling in each medicinal material sub-window and map it to the global coordinate to construct a cross-window adjacency graph composed of medicinal material joint seam continuity and shape compatibility, and generate a global Chinese medicinal material defect instance set through merging and non-maximum suppression; Step 8, generate a Chinese medicinal material slice body foreground response map, obtain a Chinese medicinal material slice body instance set through distance transformation and watershed method, and repair the joint seam contour according to the cost of the medicinal material joint seam continuity field; Step 9, construct a bipartite graph of defect instances and Chinese medicinal material slice body instances, give the edge weight by the region intersection ratio, boundary direction difference and joint seam continuity integral, and solve the maximum weight matching to obtain the Chinese medicinal material defect attribution mapping; Step 10, generate a defect statistical label according to the defect attribution mapping, write the Chinese medicinal material batch domain descriptor, mapping parameters, reference quantile group, judgment threshold and joint seam continuity summary into the Chinese medicinal material batch file, and monitor the drift in the sliding time window to trigger parameter updating. The working process of step 2 comprises:

2. The machine vision-based Chinese medicinal material defect detection method according to claim 1, characterized in that: Perform local normalization processing on the medicinal material image data in the sub-window in the spatial domain to generate a medicinal material base matrix with balanced contrast; Calculate three types of indexes related to the surface texture of Chinese medicinal materials, including gradient variance, texture energy and color entropy, based on the base matrix, and set adaptive threshold values according to quantile distribution for joint screening; In the screening process, the minimum boundary distance constraint is performed on the candidate regions to remove abnormal regions located on the extended edges of medicinal materials and stacked slices of medicinal materials; Generate a primary Chinese medicinal material anchor region set through hierarchical connected component analysis on the retained candidate regions, and perform cross-frame consistency detection on the primary anchor region set to remove transient noise spots; Output the Chinese medicinal material anchor region set after spatial uniformity constraint correction. The working process of step 3 comprises:

3. The machine vision-based Chinese medicinal material defect detection method according to claim 1, characterized in that: Convert the image to CIELab color space in the Chinese medicinal material anchor region, and generate a 32-level normalized color histogram for each channel as the Chinese medicinal material color histogram statistics; Perform two-dimensional fast Fourier transform on the anchor region grayscale image, segmentally integrate the amplitude spectrum along the radial direction, and extract the first and second order power spectrum moments to characterize the surface frequency spectrum features of Chinese medicinal materials; ​ The gray level co-occurrence matrix with a pixel pitch of 1 and directions of 0°, 45°, 90° and 135° is constructed, and four types of statistics, including contrast, correlation, energy and homogeneity, are calculated to represent the micro-texture of the Chinese medicinal material; The histogram statistics, power spectrum matrix and gray level co-occurrence features are spliced in a fixed order to form a batch domain descriptor of the Chinese medicinal material, and based on the 25%, 50% and 75% quantile of the area of the anchor region, quantile groups are divided, and the mean and covariance matrix of each quantile group are calculated as the reference Chinese medicinal material descriptor.

4. The machine vision-based Chinese medicinal material defect detection method according to claim 1, characterized in that: The working principle of step 4 is as follows: The feature vector composed of the color histogram statistics of the traditional Chinese medicinal material, the band power spectrum moment and the gray level co-occurrence matrix feature splicing is extracted for each sub-window The mean vector is calculated in the corresponding area percentile group The covariance matrix The difference score of the traditional Chinese medicinal material sub-window is calculated based on the weighted Mahalanobis distance The calculation formula is: wherein, denotes the feature vector of the sub-window, denotes the feature mean of the sub-window's area fraction group, is the inverse of the corresponding covariance matrix, is a diagonal matrix for SNR weighting of the band features, whose diagonal elements denotes the SNR estimate of the i-th band. Then, the quantile ranks of each sub-window in its four neighbors and the two neighboring frames are calculated, and these quantile ranks are fused into the original score according to the geometric average rule, and the fusion weight is guided by the SNR estimate value to reflect the reliability of each frequency band; finally, the quantile points of the reference region are counted and recorded to obtain the original score set of all sub-windows and the corresponding reference quantile points.

5. The machine vision-based Chinese medicinal material defect detection method according to claim 1, characterized in that: The score domain monotone segmentation mapping process of step 5 is as follows: The Mahalanobis distance between the batch domain descriptor of the Chinese medicinal material and the reference Chinese medicinal material descriptor is measured to obtain a quantitative difference vector; The corresponding relationship between the quantitative difference vector and the original score set of the sub-window is fitted by the B-spline piecewise interpolation algorithm to generate a segmented response function; The node position and interval slope of the segmented response function are solved by the anchor area quantile constraint to obtain a mapping parameter set and generate a mapping function; The mapping function is used for numerical transformation of all original scores to obtain a calibrated score sequence, and the iteration updates the node slope by the quantile point back substitution constraint, and stops when the mapping residual is less than the set convergence threshold; Based on the calibrated score sequence, a fixed Chinese medicinal material judgment threshold is generated, and the fixed judgment threshold is determined by the quantile position constraint method of the calibrated score distribution.

6. The machine vision-based Chinese medicinal material defect detection method according to claim 1, characterized in that: The calculation method of the structural tensor direction and the cross-seam intensity difference of step 6 is as follows: An image of the joint tape area is acquired, and the joint tape is sampled at a fixed sampling step length on both sides along the short axis A local sub-region is extracted ; The gray scale image of each local sub-region is obtained An input structure tensor calculation module calculates horizontal gradient and vertical gradient respectively, and generates a structure tensor by weighting through a Gaussian kernel function The calculation formula is: wherein, denotes a structure tensor matrix, and denote a pixel horizontal and vertical gradients, , denotes a smoothing scale, denotes a convolution operation; right Eigenvalue decomposition yields With the main direction Calculate the local consistency coefficient ,in, To prevent division by zero of small constants; Calculate intensity mean along left and right subregions With standard deviation Generate normalized cross-slit intensity difference ; Calculate the direction difference And construct the direction consistency factor ; Will , And Fusion to get the seam continuity coefficient , The formula is: wherein is a decay weight coefficient, are left and right sub-region consistency coefficients, respectively; along the discrete sampling points of the seam ridge line calculating the continuity line integral , the formula is: and based on a preset threshold value performing morphological processing to generate a boundary mask for seam reliability ranking across the window adjacency graph.

7. The machine vision-based Chinese medicinal material defect detection method according to claim 1, characterized in that: The working process of step 7 includes: In each medicinal material sub-window, the calibrated score sequence is binarized by a threshold comparison algorithm to generate a preliminary medicinal material defect mask; The preliminary defect mask is clustered by a connected domain labeling algorithm to obtain a local defect candidate region of the medicinal material; The geometric center and boundary of the local defect candidate region are mapped to the global coordinate system to obtain an initial global node set; For any adjacent sub-window defect node, the medicinal material joint continuity integral and shape compatibility index are calculated, and are used as the edge weight of the cross-window adjacency graph; The connected nodes are merged on the cross-window adjacency graph according to the edge weight threshold by the restricted union-find set algorithm to generate a preliminary global medicinal material defect instance. The preliminary global medicinal material defect instance is subjected to a weighted intersection-union ratio non-maximum suppression process, wherein the weighted intersection-union ratio index is calculated by combining the joint continuity integral and the shape compatibility; If the weighted intersection-union ratio of the candidate instance is lower than the maximum value in the neighborhood, the instance is retained and the overlapping region is removed to generate a final global medicinal material defect instance set and a corresponding instance mask. 8.The machine vision-based Chinese medicinal material defect detection method according to claim 1, characterized in that: In step 8, the working process of generating a medicinal material piece foreground response graph and performing watershed segmentation includes: The global defect instance mask and the original sub-window image are input into a foreground enhancement module to generate a medicinal material piece foreground response graph by adaptive background subtraction and normalized filtering; The distance transformation is performed on the foreground response image of the medicinal material sheet body to calculate the Euclidean distance of each foreground pixel to the nearest background pixel, and a distance mapping matrix is obtained ; based on local maximum value extraction inner point seed set , using and gradient map as a marker input watershed segmentation algorithm to generate a preliminary instance set of medicinal material pieces; for the segmentation boundary of the cross joint seam area, a seam continuity field boundary cost function , the boundary curve of the watershed segmentation is iteratively corrected, and the calculation formula of the boundary cost function is: wherein, is a sampling step size, is an index of a discrete sampling point along the watershed segmentation boundary or seam contour, is a coordinate of a discrete sampling point along the seam or segmentation boundary, is a seam continuity coefficient value corresponding to the sampling point; and finally output a set of medicinal material slice instances and a corrected boundary mask.

9. The machine vision-based Chinese medicinal material defect detection method according to claim 1, characterized in that: The working principle of the step 9 is: establishing the defect node set D and the medicinal material piece node set P in a global coordinate system; for any candidate pair With Three types of original quantities are calculated, the first type is the area intersection ratio, the second type is the weighted average of the boundary normal angle cosine, and the weight is the seam continuity coefficient at the corresponding boundary sampling point; the third type is the line integral of the seam continuity along the coincident boundary, and is normalized according to the coincident length; Constructing area ratio to expected overlap kernel regression mapping by batch level statistics and normalizing the area intersection ratio accordingly; Solving non-negative normalized weight vectors by restricted least squares based on batch anchor samples and forming a comprehensive edge weight is a linear combination of three types of normalized quantities; when the node centroid distance is less than or equal to a preset threshold and w is not lower than a dynamic threshold determined by the candidate edge weight local distribution parameter establish an edge on the bipartite graph; Performing maximum weight matching on the constructed bipartite graph to obtain the attribution mapping, and when the weight difference of the matching is not more than the discrimination threshold, comparing the seam continuity line integral first, and then comparing the boundary normal similarity to determine the attribution mapping; The home mapping and , and associated batch calibration parameter records are logged into the batch archive.

10. The machine vision-based Chinese medicinal material defect detection method according to claim 1, characterized in that: The working process of step 10 comprises: Generating a Chinese herbal medicine defect statistical label set according to the attribution mapping, wherein the label comprises defect category count, defect area proportion, centroid spatial distribution, medicine piece coincidence rate and defect continuity score; Writing the Chinese herbal medicine batch domain descriptor, score domain mapping parameter set, reference quantile group, discrimination threshold and medicine seam continuity summary into the Chinese herbal medicine batch archive according to the time stamp and version; Performing two-sample KS test, CUSUM change point detection and exponentially weighted moving average tracking on the Chinese herbal medicine anchor region quantile and calibrated score distribution within the sliding time window, and if any detection statistic exceeds the pre-set significance limit, triggering the parameter update process; the parameter update process uses the limited weighted incremental least squares fitting to solve the score domain mapping parameters, and introduces the ridge regularization term and anchor sample weight decay constraint in the fitting and updates the discrimination threshold according to the re-estimated calibrated score quantile; Writing the parameter difference before and after the update, the trigger log and the sample statistics used for testing into the Chinese herbal medicine batch archive and recording the version and rollback identifier.

Images