Processing whole-process control method based on multi-source sensing fusion

Abstract

The present application relates to the technical field of image recognition, in particular to a processing whole-process control method based on multi-source sensing fusion. The processing end face image is subjected to fuzzy restoration processing, and a restored end face image is output. The texture gradient of the longitudinal fiber is extracted from the surface texture image to generate a fiber direction vector. The probability of parallel tear is evaluated based on the spatial included angle between the fiber direction vector and a preset cutting plane, and the prior physical characteristics are obtained by conversion. The restored end face image is subjected to region growth, and the skin area of the outer ring and the xylem area of eccentric growth are separated. The hole features are identified and filtered out in the xylem area to obtain a xylem mask. Sparse sampling is performed in the xylem mask area, the cross section defect contour is located, and the edge fractal dimension feature is extracted. The fiber direction vector, the prior physical characteristics and the edge fractal dimension feature are spliced to form a multi-source feature, which is input into a morphological classification network to generate a corresponding process abnormality early warning signal and output.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, specifically to a process control method based on multi-source sensor fusion. Background Technology

[0002] Rhizome medicinal materials typically require a continuous slicing process to produce slices for subsequent processing or packaging. The quality of the slices directly affects the retention of the active ingredients and the pass rate of subsequent processing. Real-time visual inspection of the processed end face is necessary downstream of the cutting station to determine if any process abnormalities have occurred. However, rhizome medicinal materials generally exhibit biological structural characteristics such as eccentric distribution of xylem, multi-core development, and associated natural pores during their natural growth. These natural structures often appear as dark areas with low local grayscale and irregular boundaries in end face images, which are highly similar in image representation to actual cutting defects caused by tearing along the grain, moisture loss, or tool wear during the cutting process.

[0003] Existing machine vision-based inspection methods mostly rely on uniform geometric contour comparison or global threshold segmentation to process end-face images. They lack the ability to distinguish the pixel-level spatial relationship between natural pores and actual cutting defects in medicinal materials, easily misjudging natural pores as cutting defects, resulting in a high false alarm rate. In addition, after identifying end-face anomalies, existing methods can usually only provide a general judgment of whether the material is qualified or not, without further distinguishing the specific process causes of the anomaly, or tracing whether it is caused by the moisture state of the medicinal material itself, the deviation of the feed angle, or the condition of the cutting edge of the tool. This makes it difficult for on-site process personnel to obtain targeted adjustment basis.

[0004] To address this, a process control method based on multi-source sensor fusion is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a process control method based on multi-source sensor fusion, which splices the fiber orientation vector, the prior physical features and the edge fractal dimension features into a multi-source feature, inputs them into a morphology classification network, generates corresponding process anomaly early warning signals and outputs them.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A process control method based on multi-source sensor fusion includes: The surface texture image of the medicinal material before cutting and the processed end face image after cutting are obtained. The processed end face image is subjected to blur restoration processing, and the restored end face image with preserved contour details is output. The texture gradient of the longitudinal fibers is extracted from the surface texture image to generate the fiber direction vector; the probability of tearing along the grain is evaluated based on the spatial angle between the fiber direction vector and the preset cutting plane, and then converted into prior physical features. Region growing is performed on the restored end face image to separate the outer cortex region from the eccentrically grown xylem region; pore features are identified and filtered out within the xylem region to obtain a xylem mask; Sparse sampling is performed within the xylem mask area to locate the profile of the cut surface defect caused by cutting, and the edge fractal dimension features of the cut surface defect profile are extracted. The fiber orientation vector, the prior physical features, and the edge fractal dimension features are concatenated into a multi-source feature set and input into a pre-trained morphology classification network to generate and output corresponding process anomaly warning signals.

[0007] The steps for obtaining the restored end face image include: using a front-end vision acquisition device set in the feeding area to optically photograph the uncut surface of the medicinal material to obtain a surface texture image containing surface longitudinal wrinkle data; using a rear-end vision acquisition device set behind the cutting station to capture the processed end face image after cutting; extracting edge pixels from the processed end face image, calculating the gray-level gradient change rate, and extracting the spatial mapping direction with the largest gray-level gradient change rate as the edge blur direction feature; and extracting the trailing diffusion length of the edge pixels as the blur scale feature; and numerically reconstructing the edge blur direction feature and the blur scale feature to eliminate the dynamic trailing data in the processed end face image, thus obtaining the restored end face image.

[0008] The steps for obtaining prior physical features include: using the structure tensor operator to extract the directional gradient data of grooves and cracks in the surface texture image at multiple scales; performing Gaussian smoothing filtering on the directional gradient data and aggregating it to obtain the fiber direction vector characterizing the overall physical structure orientation of the medicinal material; and calculating the spatial angle data between the fiber direction vector and the normal vector of the preset cutting plane in the established image space coordinate system. The spatial angle data is input into a preset probability mapping function to obtain the longitudinal tear probability; the longitudinal tear probability is numerically compared with a preset multi-level risk threshold range to obtain the discretized risk level corresponding to the longitudinal tear probability; the discretized risk level is converted into a multi-dimensional vector format using a one-hot encoding algorithm, and feature upscaling mapping calculation is performed to generate a feature tensor with the same data dimension as the fiber direction vector, which serves as the prior physical feature.

[0009] The steps for separating the cortex and xylem regions include: calculating the local grayscale variance data of pixels in the acquired restored end-face image using a sliding window of a preset size; selecting multiple pixel nodes whose local grayscale variance data is greater than the peak variance threshold and converging them into an initial seed point set to adapt to the multi-core biological development characteristics of the eccentric growth of medicinal materials; starting from each pixel node in the initial seed point set, performing adaptive region growth operations on the outer pixels according to a preset texture similarity threshold, and simultaneously updating the region attribution label data of adjacent pixels; until the local regions grown from different initial seed points merge morphologically and stop the growth operation when they touch the grayscale gradient closed boundary, obtaining the fused growth region data; marking the fused growth region data as the eccentrically grown xylem region; and marking the range data outside the closed boundary up to the edge of the restored end-face image as the outer cortex region, thus completing the separation.

[0010] The steps for obtaining the xylem mask include: performing a global spatial search of low-grayscale connected dark area data within the marked xylem region to obtain an initial set of dark areas; when the grayscale mean of each connected dark area in the initial set of dark areas meets the environmental background threshold condition, tracking and calculating the boundary extension trajectory data of the connected dark areas; determining whether the boundary extension trajectory data is completely enclosed within the pixel coordinates of the xylem region and does not have any spatial intersection with the surrounding cortex region; if the determination result is yes, then the connected dark areas are determined to be pore features formed by natural physiological development inside the medicinal material; generating an initial binarization matrix with the same size as the restored end face image, filtering out the matrix elements corresponding to the pore features, and outputting the processed initial binarization matrix as the xylem mask.

[0011] The steps for extracting the edge fractal dimension features of the cut surface defect contour include: generating a sparse sampling grid matrix within the effective pixel coordinate range constrained by the xylem mask according to a preset sampling step size; extracting only the pixel grayscale data located at the intersection nodes of the sparse sampling grid matrix, and calculating macroscopic texture statistics data including grayscale skewness and kurtosis parameters. By comparing macroscopic texture statistics with preset end-face standard data, boundary regions with discontinuous gray-level abrupt changes are located, and the set of connected edge pixels is extracted as the cut surface defect contour. The box-counting dimension algorithm is used to analyze the topological features of the cut surface defect contour. Multiple square grids with continuously decreasing side lengths are used to cover the cut surface defect contour, and the number of grids required for complete coverage each time is counted. In the logarithmic coordinate system, the linear slope value between the reciprocal of the grid side length and the number of grids is fitted. The fitted linear slope value is used as the edge fractal dimension feature characterizing the degree of edge roughness and fracture.

[0012] The steps for generating the corresponding process anomaly warning signal include: performing extreme value normalization on the extracted fiber direction vector and edge fractal dimension features respectively; combining the normalized fiber direction vector and edge fractal dimension features with prior physical features, and splicing and recombining them according to the preset tensor channel dimension to construct a multi-source feature tensor with unified dimension; inputting the multi-source feature tensor into the morphological classification network for nonlinear mapping and network weight calculation; outputting the probability distribution value of each category in which the current image state belongs to the categories of excessive moisture loss, tearing under stress during cutting along the grain, and blunting of the cutting edge; selecting the category with the highest probability distribution value as the final morphological classification result, and generating a process anomaly warning signal bound to the morphological classification result for terminal output.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention addresses the biological structural characteristics of the xylem in rhizomatous medicinal materials, which exhibit eccentric growth and multi-core development. First, it spatially separates the xylem and cortex regions. Then, it performs grayscale background thresholding and spatial topological determination of boundary extension trajectories for connected dark areas within the xylem region. Since natural pores in medicinal materials are typically completely surrounded by the xylem in space, this processing reduces false alarms when identifying natural pores solely based on grayscale or shape, improves the accuracy of identifying actual cutting defects, and provides a reliable data foundation for subsequent extraction of defect contours.

[0014] 2. Within the effective pixel coordinate range defined by the xylem mask, this invention sets a sparse sampling grid according to a preset step size. The grayscale data at the sparse sampling nodes are used to calculate macroscopic texture statistics such as grayscale skewness and kurtosis, replacing the point-by-point traversal of all pixels. When the macroscopic texture statistics indicate that there is a discontinuous change in grayscale in a local area, the cross-sectional defect contour is extracted from the corresponding area, and box-counting dimension calculation is performed. By combining large-scale coarse-grained statistics with fine-grained topological analysis of local areas, the accuracy of defect contour recognition is maintained while reducing the amount of data processing in the system and reducing the possibility of detection lag due to excessive computation.

[0015] 3. This invention unifies and normalizes the fiber direction vector transformed from the surface texture before cutting, the prior physical features of tearing along the grain, and the edge fractal dimension features transformed from the defect contour of the cut rear face, and splices them according to the channel dimension to construct a multi-source feature tensor. This tensor is then input into a morphological classification network, and outputs the probability distributions corresponding to three specific causes: excessive moisture loss, tearing under stress during cutting along the grain, and blunting of the tool edge. This generates a process anomaly warning signal bound to the cause category, providing a clear basis for targeted adjustment of lubrication process, feed angle, or tool maintenance. Attached Figure Description

[0016] Figure 1This is a flowchart illustrating the process control method for the entire manufacturing process based on multi-source sensor fusion according to the present invention. Figure 2 This is a logical schematic diagram of the processing whole-process control method based on multi-source sensor fusion of the present invention; Figure 3 This is a schematic diagram of the morphological classification network of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example 1

[0018] This invention proposes a process control method based on multi-source sensor fusion for the entire processing workflow. This method is applied to edge vision image early warning during the slicing process of root and rhizome medicinal materials. The process flow is as follows: Figure 1 As shown, the data logic of the method is as follows: Figure 2 As shown, it includes: The surface texture image of the medicinal material before cutting and the processed end face image after cutting are obtained. The processed end face image is subjected to blur restoration processing, and the restored end face image with preserved contour details is output. The texture gradient of the longitudinal fibers is extracted from the surface texture image to generate the fiber direction vector; the probability of tearing along the grain is evaluated based on the spatial angle between the fiber direction vector and the preset cutting plane, and then converted into prior physical features. Region growing is performed on the restored end face image to separate the outer cortex region from the eccentrically grown xylem region; pore features are identified and filtered out within the xylem region to obtain a xylem mask; Sparse sampling is performed within the xylem mask area to locate the profile of the cut surface defect caused by cutting, and the edge fractal dimension features of the cut surface defect profile are extracted. The fiber orientation vector, the prior physical features, and the edge fractal dimension features are concatenated into a multi-source feature set and input into a pre-trained morphology classification network to generate and output corresponding process anomaly warning signals.

[0019] Preferably, the step of acquiring the restored end face image includes: using a front-end vision acquisition device set in the feeding area to optically photograph the uncut surface of the medicinal material to acquire a surface texture image containing surface longitudinal wrinkle data; using a rear-end vision acquisition device set behind the cutting station to capture the processed end face image after cutting; extracting edge pixels in the processed end face image, calculating the gray-level gradient change rate, and extracting the spatial mapping direction with the largest gray-level gradient change rate as the edge blur direction feature; and extracting the trailing diffusion length of the edge pixels as the blur scale feature; and numerically reconstructing the edge blur direction feature and the blur scale feature to eliminate the dynamic trailing data in the processed end face image, thus obtaining the restored end face image.

[0020] In this embodiment, the multi-source sensor data is provided by a front-end vision acquisition device and a back-end vision acquisition device. Both devices are synchronously triggered by a conveyor encoder, and the same number is assigned to the same herb according to the time difference between the herb's passage through two shooting positions, ensuring a one-to-one correspondence between the images before and after cutting. The camera's internal and external parameters are determined by a calibration plate; the herb's posture is constrained by a guiding mechanism. Under the condition that the surface observation area can be approximated as a local plane, the texture direction in the image is mapped to a device coordinate system established with the conveying direction and the normal direction of the cutting plane.

[0021] In the specific implementation process, an industrial area array camera is installed above the feeding station of the medicinal herb slicing processing line as a front-end visual acquisition device. Its shooting direction is towards the surface of the medicinal herb that is about to enter the cutting station, so as to perform a static optical image of the outer skin of the medicinal herb before cutting and obtain a surface texture image. Since the medicinal herb is still in a static or uniformly conveyed state at this time, and the object of the shooting is the skin, the obtained image clearly preserves the texture details such as wrinkles, grooves and cracks extending longitudinally on the surface of the medicinal herb. This image serves as the original input for the subsequent prior physical feature extraction step.

[0022] Downstream of the cutting station, another industrial camera is installed along the conveyor belt's forward direction as a back-end vision acquisition device. Its field of view is aimed at the processed end face exposed after the medicinal material has been cut. As the processed end face moves continuously on the conveyor belt at a certain speed, the pixels will shift along the conveyor direction during the camera's exposure time, thus forming a trailing shadow in the image that extends along the conveyor direction. This trailing shadow will obscure the original high-frequency details of the end face itself, such as the fine serrated edges caused by fiber breakage.

[0023] To estimate the motion blur direction, edge detection is first performed on the processed end face image to obtain edge pixels with abrupt gray-level changes. The gray-level gradient change rate along a certain direction can be characterized by the rate at which the gray-level gradient amplitude decays with distance from the edge: the slower the gray-level gradient amplitude decays and the wider the influence range in that direction, the greater the gray-level gradient change rate in that direction. For each edge pixel, candidate directions with equal angles are set within the range of 0 to 180 degrees. The diffusion width of the gray-level gradient from its peak to the background level is statistically analyzed along each candidate direction, and this diffusion width is used as the specific characterization value of the gray-level gradient change rate for the corresponding candidate direction. A robust average of the diffusion widths of all edge pixels is performed, and the direction with the largest average diffusion width, i.e., the largest gray-level gradient change rate, is determined as the edge blur direction feature. This criterion focuses on the edge diffusion width and does not directly equate the original gray-level gradient direction with the object's motion direction.

[0024] After determining the edge blur direction features, for each edge pixel, the gray values ​​of adjacent pixels are sampled point by point along this direction. The number of pixels the gray value passes through from the edge to the background level is counted. This number of pixels is converted into the actual physical length or pixel length and used as the trailing diffusion length at the edge pixel. The average or weighted average length of the trailing shadows of all edge pixels is taken to obtain a blur scale feature that represents the degree of trailing shadows in the entire image. After obtaining the edge blur direction feature and blur scale feature, a motion blur kernel corresponding to the current trailing shadow is constructed using the angle represented by the edge blur direction feature and the length represented by the blur scale feature as parameters. Mathematically, the motion blur kernel is represented by a one-dimensional linear distribution function along the edge blur direction with a length comparable to the blur scale feature. Subsequently, the original processing end face image is deconvolved using this motion blur kernel. That is, in the frequency domain or spatial domain, the image is inversely filtered using this motion blur kernel to eliminate the dynamic trailing shadow data caused by the conveyor belt movement, restore the edge details originally obscured by the trailing shadows in the end face image, and obtain a restored end face image that retains the contour details. The deconvolution uses Wiener filtering or inverse filtering with regularization terms. The noise power is estimated from the background area outside the end face to avoid directly inverting and amplifying high-frequency noise. The restoration parameters are selected based on improving edge sharpness without obvious ringing. This invention avoids the detection limitations caused by incomplete image information from a single viewpoint by setting up separate front-end and back-end visual acquisition. At the same time, by extracting the directional and scale features of edge blurring and performing numerical reconstruction accordingly, it can effectively eliminate the dynamic trailing generated during high-speed assembly line movement, restore the high-frequency edge details that the processed end face should have, and provide clear and reliable basic image data for subsequent steps such as region separation, hole filtering, and extraction of cross-sectional defect contours, thereby improving the image quality of the overall detection process and the accuracy of subsequent feature extraction.

[0025] Preferably, the steps for obtaining prior physical features specifically include: In the specific implementation process, the obtained surface texture image is first processed using a structure tensor operator. The specific operation method of the structure tensor operator is as follows: for each pixel in the image, its first-order gray-level gradient components in the horizontal and vertical directions are calculated, and a second-order symmetric matrix is ​​constructed using these two directional components. This matrix is ​​the structure tensor at that pixel. To take into account the two different scales of texture features in the image, namely coarse grooves and fine cracks, the above structure tensor calculation process is repeated under multiple preset neighborhood windows of different sizes, thereby obtaining the set of structure tensors at multiple scales for that pixel. Eigenvalue decomposition is performed on the structure tensor at each scale, and the direction indicated by the eigenvector corresponding to its smallest eigenvalue is taken as the directional gradient data of the groove or crack direction at that pixel at that scale; this direction is orthogonal to the direction of the most drastic gray-level change (i.e., the direction corresponding to the largest eigenvalue). Subsequently, the directional gradient data obtained at each scale were processed using Gaussian smoothing filtering. Specifically, with each pixel as the center, the directional gradient data of each pixel in its neighborhood was weighted and averaged according to the weights determined by the Gaussian function. This eliminated the interference of local noise points on the direction judgment, resulting in a smoothed directional gradient field. This yielded a unified directional numerical vector that comprehensively represents the longitudinal texture of the overall epidermis of the medicinal material; this vector is the fiber direction vector. Since the fiber direction is an axial vector rather than a directed vector, the directional gradient data at each scale were converted to a double-angle representation before weighted aggregation. After aggregation, the resulting angle was divided by two and mapped back to the original angle range to avoid distortion of the aggregated direction or low amplitude due to direction cancellation. Mathematically, the fiber direction vector is a direction vector with a definite angular component, representing the dominant direction of the epidermal texture of this batch of medicinal materials extending along the longitudinal axis of the medicinal material.

[0026] After obtaining the fiber direction vector, it is mapped to the equipment coordinate system established with the cutting station as the reference, based on the camera calibration results. The normal vector of the cutting plane is then determined according to the tool mounting position and cutting posture. The fiber direction vector and the normal vector of the cutting plane are normalized to obtain the unit fiber direction vector and unit normal vector, respectively. Since the texture of the medicinal material fibers only represents the axial direction and does not distinguish between positive and negative directions, the absolute value of the dot product of the unit fiber direction vector and the unit normal vector is calculated, and the inverse cosine of this absolute value is taken to obtain the acute angle between the fiber direction and the normal direction of the cutting plane. ,Right now:

[0027] in, Represents the unit fiber direction vector; Represents the unit normal vector; The value ranges from 0 to 90 degrees; further, based on 90 degrees minus Calculate the deviation angle between the fiber direction and the cutting plane. Deviance angle When the value is 0, it indicates that the fiber direction is located or nearly parallel to the cutting plane; deviation angle An increase in the angle indicates that the fiber direction is gradually deviating from the cutting plane. The deviation angle is a risk mapping parameter calculated based on the spatial angle between the fiber direction vector and the normal vector of the cutting plane, used to characterize the relative relationship between the fiber orientation of the medicinal material and the current cutting posture.

[0028] The specific relationship between the deviation angle and the risk of tearing along the grain is determined based on historical cutting samples under the same medicinal material varieties and processing equipment conditions, rather than directly interpreting the angle between the fiber direction and the normal direction as the angle between the fiber direction and the cutting plane. Furthermore, the deviation angle is input into a preset probability mapping function. As one implementation method, a binomial logistic regression model from the prior art is used to establish the mapping relationship between the deviation angle and the probability of tearing along the grain. The probability of tearing along the grain is calculated according to the following formula:

[0029] in, This represents the probability of tearing along the grain, with a value ranging from 0 to 1; This represents a kurtosis coefficient greater than 0, used to control how quickly the probability changes with the deviation angle; Indicates the position parameter, when equal At that time, the probability of tearing along the grain is one in half. and The parameters are obtained by fitting historical processing samples labeled with actual tearing results. During fitting, samples that actually exhibited tearing along the grain are labeled as 1, and samples that did not exhibit tearing along the grain are labeled as 0. The parameters are determined using either maximum likelihood estimation via logistic regression or minimum cross-entropy loss via binary cross-entropy. The deviation angle and position parameters use the same angular unit; when both are expressed in degrees, the kurtosis coefficient is measured per degree; when both are expressed in radians, the kurtosis coefficient is measured per radian, so that the exponential term is dimensionless. The same angular unit is used for model training and actual calculations; a decrease in the deviation angle increases the probability of tearing along the grain, and an increase in the deviation angle decreases the probability of tearing along the grain. For cases where the actual samples do not satisfy a monotonic relationship, a piecewise linear probability mapping function can also be used based on historical samples, but the continuity between each segment should be ensured, and the output value should be limited to the range of 0 to 1.

[0030] A first probability threshold and a second probability threshold greater than the first probability threshold are preset. When the probability of tearing along the grain is less than the first probability threshold, it is determined to be a low-risk level; when the probability of tearing along the grain is greater than or equal to the first probability threshold and less than the second probability threshold, it is determined to be a medium-risk level; when the probability of tearing along the grain is greater than or equal to the second probability threshold, it is determined to be a high-risk level. For the first and second probability thresholds: the candidate probability thresholds are gradually changed on independent validation samples, and the false positive rate and false negative rate are calculated respectively; the lower probability threshold corresponding to the false negative rate not exceeding the preset upper limit is determined as the first probability threshold, and the higher probability threshold corresponding to the false positive rate not exceeding the preset upper limit is determined as the second probability threshold. The second probability threshold is set to ensure that low-risk, medium-risk, and high-risk levels meet the preset risk differentiation requirements of the production line. Finally, the discretized risk level is converted into a multi-dimensional vector using a one-hot encoding algorithm, where the position corresponding to the current risk level is set to 1, and the other positions are set to 0. To ensure that the encoded vector is consistent with the fiber direction vector in terms of data dimension, so as to facilitate subsequent unified feature splicing operations, the feature upscaling mapping calculation is further performed on the one-hot encoded vector. That is, through a pre-trained linear mapping matrix, the dimension of the one-hot encoded vector is expanded to the same dimension value as the fiber direction vector, thereby obtaining the final prior physical feature tensor.

[0031] This invention abstracts the longitudinal texture of medicinal herb epidermis into a quantifiable fiber direction vector by using multi-scale structural tensors and Gaussian smoothing aggregation. Combined with the spatial angle relationship between the normal vectors of the pre-defined cutting plane, the probability of tearing along the grain is derived. Furthermore, through discretization and hierarchical classification, one-hot encoding, and feature upscaling mapping, this probability scalar is transformed into a high-dimensional feature tensor with the same dimension as the fiber direction vector. This avoids the problem of single probability values ​​being diluted in weight and information being submerged in the subsequent multi-source feature fusion process, and provides a physically meaningful prior basis for subsequent comprehensive classification.

[0032] Preferably, the step of separating the cortex region and the xylem region includes: calculating the local gray-level variance data of pixels in the acquired restored end face image using a sliding window of a preset size; selecting multiple pixel nodes whose local gray-level variance data is greater than the peak variance threshold and converging them into an initial seed point set to adapt to the multi-core biological development characteristics of the eccentric growth of medicinal materials; starting from each pixel node in the initial seed point set, performing adaptive region growth operation on the outer pixels according to a preset texture similarity threshold, and synchronously updating the region attribution label data of adjacent pixels; until the local regions grown from different initial seed points merge morphologically and stop the growth operation when they touch the gray-level gradient closed boundary, obtaining fused growth region data; marking the fused growth region data as the eccentrically grown xylem region; and marking the range data outside the closed boundary up to the edge of the restored end face image as the outer cortex region, thus completing the separation.

[0033] For the restored end face image, a square sliding window of fixed size is set, and the sliding window moves row by row and column by column according to a preset step size. Whenever the sliding window reaches a pixel position, the gray value of each pixel within the window coverage area is counted, the corresponding gray value variance is calculated, and the obtained variance is assigned to the center pixel of the sliding window. After traversing the entire image, a local gray value variance data matrix corresponding to the size of the restored end face image is formed. The xylem of rhizome medicinal materials may have multiple eccentrically distributed growth cores. The tissue texture near the growth cores is relatively concentrated, and the gray value fluctuation is usually more obvious, so it is represented by a relatively high variance value in the local gray value variance data matrix.

[0034] Based on a pre-set peak variance threshold, pixels with variance values ​​exceeding the threshold are selected from the local grayscale variance data matrix, and these selected pixels are used as the initial seed point set. The initial seed points can be distributed in multiple separate locations within the xylem to correspond to its multi-core and eccentric growth tissue characteristics. During region growth, each initial seed point is used as a starting point to sequentially examine adjacent pixels near the current region boundary and compare their grayscale values ​​or texture features. When the comparison result meets the preset texture similarity condition, the adjacent pixel is merged into the corresponding region, and the region's attribution label is updated. If the condition is not met, the pixel is not merged. The above process proceeds outward layer by layer in a breadth-first manner. When local regions formed by different seed points come into contact with each other, if the boundary pixels at the contact points still meet the texture similarity condition, the corresponding regions are merged and labeled uniformly. After the region growth front reaches a boundary with obvious gradient changes and continuous closure, it stops expanding along that direction; until all directions no longer meet the growth conditions, the fused growth region data is output. The grayscale gradient closure boundary is located at the junction of xylem and cortex, and consists of pixels whose gradient magnitude is continuously higher than a preset threshold and forms a closed curve. First, the gradient magnitude of the restored end face image is calculated using the Sobel operator. The 10th percentile of the gradient magnitude at the junction of normal samples is set as the gradient threshold. Then, morphological closing operations are used to connect boundary discontinuities with a length of less than three pixels. When the area enclosed by the closed curve is within the area of ​​normal xylem, the curve is used as the stopping boundary for region growth. If the region growth process continues until the preset maximum number of iterations, or if the area of ​​the growing region exceeds the upper limit of the normal xylem area and a boundary that meets the closure condition is still not formed, then growth is stopped and the current sample is marked as having insufficient confidence in tissue separation. The corresponding verification mark is output, and the subsequent hole identification and defect contour extraction steps are no longer executed. The fused growth region is marked as an eccentrically growing xylem region; the area between the outer edge of the boundary and the outer edge of the restored end face image is marked as the cortex region; thus, the spatial separation of the two types of tissues is completed, and the resulting xylem region marking result is sent to the subsequent hole feature recognition; the present invention achieves accurate separation of the xylem region and the cortex region through adaptive growth under texture similarity constraints and the stopping condition of gray-level gradient closed boundary, providing a clear regional range constraint for the subsequent hole feature recognition, and also eliminating the interference of cortex region texture for the extraction of cross-sectional defect contours, thereby improving the regional positioning accuracy of the overall detection process.

[0035] Preferably, the steps for obtaining the xylem mask include: In the specific implementation process, for the marked xylem region, a global spatial search is performed within the pixel coordinate range of the region to find connected pixel sets with gray values ​​significantly lower than those of the surrounding region. Each found connected pixel set is recorded as a connected dark region, and all connected dark regions together constitute an initial dark region set. For each connected dark region in the initial dark region set, the average gray value of all pixels within it is calculated, and this average value is compared with a pre-set environmental background threshold condition. Here, the environmental background threshold condition means that the average gray value of the connected dark region should be within a numerical range that conforms to the background environmental noise or the natural hollow structure inside the medicinal material. If the average gray value of the connected dark region meets this condition, the boundary pixels of the connected dark region are tracked sequentially, and the trajectory of the boundary pixel coordinates changing with the tracking direction during the tracking process is recorded to obtain the boundary extension trajectory data of the connected dark region. The environmental background threshold was determined by artificially confirmed natural hole samples from normal medicinal materials in the same batch. The mean of the gray values ​​of the holes plus or minus two standard deviations was taken as the allowable range. Before shooting, the brightness was normalized using a grayscale plate to reduce the impact of light changes on the threshold.

[0036] Furthermore, the pixel coordinates contained in the boundary extension trajectory data are compared with the pixel coordinate ranges of the xylem region and the outer cortex region to determine whether the boundary extension trajectory data falls entirely within the pixel coordinates of the xylem region and whether the trajectory overlaps or intersects with the pixel coordinates of the cortex region. If the determination result is that the boundary extension trajectory is completely surrounded by the xylem region and does not intersect with the cortex region, it indicates that the connected dark area is a hollow structure completely wrapped by the xylem tissue, which is consistent with the physiological characteristics of hollow pith or natural fissures formed in medicinal materials during natural growth or drying. The connected dark area is then identified as a cavity feature. Conversely, if the boundary extension trajectory of the dark area intersects with the cortex region, or if the dark area is not completely wrapped by the xylem region, it is not identified as a cavity feature but should be retained as a potential cutting defect for subsequent processing.

[0037] After identifying all connected dark areas, an initial binarized matrix of the same size as the restored end-face image is generated. In this matrix, the values ​​corresponding to xylem regions are set to one, and the values ​​corresponding to non-xylem regions are set to zero. For pixel coordinate ranges identified as hole features, the values ​​at the corresponding positions in this initial binarized matrix are masked and set to zero, indicating that these regions are not considered valid xylem material. The binarized matrix after this masking process is output as a xylem mask. This mask retains only a one-value marker at the valid pixel coordinate positions corresponding to xylem material, constraining subsequent processing to be performed only within the valid material range.

[0038] This invention uses background threshold determination based on the average grayscale value of connected dark areas within the xylem region, combined with topological determination of the spatial enclosing relationship between the boundary extension trajectory and the xylem and cortex regions. This allows for accurate differentiation between naturally formed physiological pores within medicinal materials and damage defects caused by external mechanical cutting, avoiding false alarms caused by general detection algorithms misclassifying natural pores as cutting defects. The generated xylem mask retains only the pixel range corresponding to the effective material, eliminating interference from natural pores for subsequent extraction of the cut surface defect contour, thus improving the accuracy of defect identification.

[0039] Preferably, the step of extracting the edge fractal dimension features includes: generating a sparse sampling grid matrix within the effective pixel coordinate range constrained by the xylem mask according to a preset sampling step size; extracting only the pixel grayscale data located at the intersection nodes of the sparse sampling grid matrix, and calculating macroscopic texture statistics including grayscale skewness and kurtosis parameters; comparing the macroscopic texture statistics with preset end-face standard data to locate boundary regions with discontinuous grayscale changes, and extracting the set of connected edge pixels as the cut surface defect contour; performing topological feature analysis on the cut surface defect contour using a box-counting dimension algorithm, covering the cut surface defect contour with multiple square grids with continuously decreasing side lengths, and counting the number of grids required for complete coverage each time; fitting the linear slope value between the reciprocal of the grid side length and the number of grids in a logarithmic coordinate system; and using the fitted linear slope value as the edge fractal dimension feature characterizing the degree of edge roughness and fracture.

[0040] In the specific implementation process, for the effective pixel coordinate range constrained by the output wood mask, grid lines are laid out at fixed intervals along the horizontal and vertical directions within this range according to a pre-set sampling step size, thus forming a sparse sampling grid matrix. The intersection points of adjacent grid lines in this grid matrix are the sparse sampling nodes. Unlike the method of calculating each pixel one by one, this step only extracts the pixel grayscale data located at the above-mentioned sparse sampling node positions, thereby significantly reducing the amount of data to be processed. Macro-texture statistics establish a local window centered on each sparse sampling node, and calculate the skewness and kurtosis of the sampled grayscale within the window; only windows exceeding the standard end face range are restored to full pixel processing, and connected edges are extracted within the window, so that the macro-statistical results can correspond to specific anomaly locations.

[0041] For the extracted sparse sampling node grayscale data set, the skewness parameter and kurtosis parameter of its grayscale distribution are calculated respectively. The skewness parameter is used to characterize the degree of deviation of the grayscale distribution from the symmetrical distribution, and the kurtosis parameter is used to characterize the sharpness or flatness of the grayscale distribution from the normal distribution. The above skewness and kurtosis parameters are used together as macroscopic texture statistics that can summarize the overall texture characteristics of the region.

[0042] Subsequently, the aforementioned macroscopic texture statistics were compared with pre-defined end-face standard data, which refers to a pre-defined range of standard values ​​based on the skewness and kurtosis parameters that the end face of the wood should exhibit under normal cutting conditions. By comparing the macroscopic texture statistics region by region to see if they exceed this standard value range, local boundary regions with discontinuous abrupt changes in grayscale were located, i.e., regions where the macroscopic texture statistics deviate significantly from the standard range. Connected component extraction was performed on the pixels in these regions to obtain a set of connected edge pixels. This set of connected edge pixels was determined as the cut surface defect contour, which corresponds to the tearing or breakage marks generated within the effective material range of the wood during the actual cutting process.

[0043] Furthermore, for the aforementioned cut-out contour, a box-counting dimension algorithm is used for topological feature analysis. Specifically, multiple square grid side lengths are pre-defined, decreasing in size from largest to smallest. For each grid side length, the square grid of that length is used to cover the entire coordinate plane of the image. The number of grids that intersect with or contain at least one pixel of the cut-out contour is counted. Since only the aforementioned grids actually cover the pixel range of the cut-out contour, and the other grids have no intersection with the contour, the aforementioned number of grids is the minimum number of grids required to completely cover the cut-out contour at that grid side length, thus obtaining the grid count data corresponding to that grid side length. The above statistical process is repeated for multiple preset different grid side lengths to obtain a set of data pairs of grid side lengths and their corresponding grid counts.

[0044] The reciprocal of the grid side length and the logarithm of the number of grids in the above data pairs are taken. In a logarithmic coordinate system with the logarithm of the reciprocal of the grid side length as the x-axis and the logarithm of the number of grids as the y-axis, linear fitting is performed on the above multiple sets of data points to obtain the slope value of the fitted line. This slope value is used as the edge fractal dimension feature. The larger the value, the higher the boundary complexity of the cut surface defect contour at different scales, that is, the rougher and more irregular the fracture. The smaller the value, the relatively smooth the fracture. This edge fractal dimension feature is finally passed as an independent type of feature data to the subsequent multi-source feature stitching stage. The box scale is at least five levels, the minimum side length is not less than two pixels, and the maximum side length does not exceed one-quarter of the short side of the circumscribed rectangle of the defect. When the determination coefficient of the double logarithmic linear fitting is lower than 0.90, the contour is re-extracted or the scale range is adjusted, and the edge fractal dimension is not directly output. This invention uses a box-counting dimension algorithm to perform topological analysis on the profile of a cut surface defect, which yields edge fractal dimension features that can quantitatively characterize the roughness and irregularity of the cut surface. Compared with a simple geometric circumscribed contour comparison method, it has a stronger ability to distinguish the fracture morphology and provides a physically meaningful topological feature basis for subsequent multi-source feature fusion and classification.

[0045] Furthermore, the steps for generating the corresponding process anomaly warning signal include: performing extreme value normalization on the extracted fiber direction vector and edge fractal dimension features respectively; combining the normalized fiber direction vector and edge fractal dimension features with prior physical features, and splicing and recombining them according to the preset tensor channel dimension to construct a multi-source feature tensor with unified dimension; inputting the multi-source feature tensor into the morphological classification network for nonlinear mapping and network weight calculation; outputting the probability distribution value of each category in which the current image state belongs to the category of excessive moisture loss, the category of tearing under stress during cutting along the grain, and the category of blunting of the cutting edge; selecting the category with the highest probability distribution value as the final morphological classification result, and generating a process anomaly warning signal bound to the morphological classification result for terminal output.

[0046] In the specific implementation process, extreme value normalization processing is performed on the numerical range of fiber orientation vector, prior physical features, and edge fractal dimension features, respectively. That is, the maximum and minimum values ​​of each type of feature in the pre-statistical historical data set are found, the current feature value is subtracted from the minimum value and then divided by the difference between the maximum and minimum values, so that the features are mapped to the same numerical range, eliminating the adverse effects on subsequent fusion caused by the differences in the original dimensions and numerical ranges of different features.

[0047] After normalization, the normalized fiber orientation vector data, prior physical feature data, and normalized edge fractal dimension feature data are sequentially concatenated and arranged at corresponding channel positions according to a pre-defined tensor channel dimension order. This constructs a multi-source feature tensor where the dimensions of various feature data are unified and each feature occupies an independent channel position. This tensor structurally preserves the numerical information of each of the three types of features and facilitates unified processing by the subsequent network. The multi-source feature tensor is then input into a pre-trained morphology classification network, which consists of a multi-layer neural network structure. The structure of the network model is as follows: Figure 3 As shown, the network specifically includes a data input layer, a first neural network layer, a second neural network layer, a third neural network layer, and a network output layer. Through layer-by-layer nonlinear mapping operations and network weight calculations, the input multi-source feature tensor is comprehensively analyzed and processed. Finally, in the network output layer, the probability value of the current input image state belonging to one of three preset categories is calculated: excessive moisture loss, tearing under stress during cutting along the grain, and blunting of the cutting edge. The sum of the probability values ​​of the three categories is one, which together constitutes the probability distribution value.

[0048] The morphological classification network employs a three-layer fully connected network as one implementation method, with 64, 32, and 16 nodes in the three layers, respectively. The hidden layers use a linear rectified function, the output layer uses a normalized exponential function, and the loss function is cross-entropy. Training samples consist of images of the same medicinal material before and after cutting. Over-drying labels are determined by measured moisture content below the process lower limit; tear-along-grain labels are determined by the relationship between the cutting plane and fiber direction, as well as manual verification; and tool dulling labels are determined by the cumulative number of cuts, edge wear measurement, and end-face verification. Samples are divided into training, validation, and test sets according to batches of medicinal materials to avoid repetition of images of the same medicinal material across sets.

[0049] After obtaining the probability distribution values ​​of the three categories, the values ​​of the three categories are compared, and the category with the highest value is selected as the final morphological classification result determined in this test. Based on the morphological classification result, the corresponding process abnormality warning signal is retrieved from the pre-set signal library. This signal contains a specific classification result identifier, such as indicating that the abnormality is caused by excessive dryness of the end face due to moisture loss of the medicinal material, or by tearing along the grain due to the cutting angle deviating from the fiber direction, or by a decrease in cutting quality due to wear and dulling of the tool edge. The process abnormality warning signal is output through the terminal equipment, and the processing parameters are adjusted or the tool replacement and maintenance are arranged in a targeted manner. Process anomaly warning signals can be transmitted to a programmable logic controller (PLC) or human-machine interface (HMI) to trigger lubrication checks, feed attitude adjustments, or tool maintenance commands. Specific adjustments are executed by the existing controller according to the upper and lower limits of the on-site process. Other preset parameters are determined as follows: the sliding window side length is five to nine pixels, selected based on the average period of the surface texture; the multi-scale structure tensor window is 3, 5, and 9 pixels, with the Gaussian standard deviation approximately one-sixth of the window side length; the sparse sampling step size is no greater than half the minimum defect width required to be detected on-site; the end-face standard data consists of the mean skewness and kurtosis of each local window in the normal cutting sample plus or minus two times the standard deviation; the maximum and minimum values ​​of the extreme value normalization are only statistically derived from the training set; the tensor channel sequence is fixed as fiber direction vector, prior physical features, and edge fractal dimension features; other parameters without fixed values ​​are determined through calibration samples, provided that the camera, light source, conveyor speed, and medicinal material variety remain consistent; when the equipment installation location, imaging conditions, conveyor speed, or medicinal material variety changes, parameter calibration is performed again.

[0050] This invention constructs a multi-source feature tensor with complete information and unified structure, and combines it with the nonlinear mapping capability of a morphological classification network to achieve deep fusion analysis of prior information on cutting posture and fracture topology. Compared with methods that rely solely on single-dimensional features for judgment, this invention can indicate different causes such as moisture loss, tearing along the grain, or tool wear, providing clear guidance for adjusting on-site process parameters. Example 2

[0051] Based on Example 1, the process of generating fiber direction vector, separating cortex region and xylem region, obtaining xylem mask, and generating process abnormality warning signal is further defined.

[0052] The step of generating the fiber orientation vector further includes: calculating the overall grayscale mean data of the surface texture image; comparing the overall grayscale mean data with a preset moisture content grayscale comparison range to determine the moisture content level corresponding to the current medicinal material; selecting a set of scale weight coefficients corresponding to the moisture content level from a preset set of scale weight coefficients according to the moisture content level; and using the selected scale weight coefficients to weight and aggregate the orientation gradient data at multiple scales to obtain the fiber orientation vector.

[0053] Beforehand, image acquisition and experimental statistics were conducted on medicinal material samples with different moisture contents to establish a comparison table between the overall grayscale mean and moisture content levels. A set of suitable scale weighting coefficients was then assigned to each moisture content level. In actual testing, the overall grayscale mean of the current surface texture image was first calculated. Based on this, the moisture content level was determined by referring to the table. The corresponding scale weighting coefficients were then retrieved to weight and aggregate the multi-scale directional gradient data. This allowed for the stable extraction of fiber direction vectors by using suitable weight combinations for different batches of medicinal materials where high moisture content resulted in weak texture contrast or low moisture content resulted in excessive texture contrast. When establishing the grayscale control interval for moisture content, a grayscale plate is placed under a fixed light source to correct the image brightness, and the moisture content of the sample is measured using the drying and weighing method. The samples are grouped according to the moisture content process interval, and the confidence interval of the overall grayscale mean after correction is calculated. The scale weight of each group is based on minimizing the fiber orientation labeling error. A grid search is performed on the verification samples, and the weights are normalized to a sum equal to 1. This invention solves the specific problem that the difference in moisture content of different batches of medicinal materials leads to inconsistent texture contrast, which in turn affects the stability of fiber orientation vector extraction, and improves the consistency of prior physical features across different batches.

[0054] The steps for separating the cortex and xylem regions further include: statistically analyzing the global grayscale histogram data of the restored end-face image; dynamically calculating the variance peak threshold and texture similarity threshold that match the current image based on the distribution peak and dispersion of the global grayscale histogram data; filtering the initial seed point set using the dynamically calculated variance peak threshold, and performing adaptive region growth operation using the dynamically calculated texture similarity threshold; specifically, statistically analyzing the global grayscale histogram of the restored end-face image and determining the main peaks of the xylem and cortex and the valleys between the two peaks; calculating the half-peak width and corresponding grayscale intra-class variance of the two main peaks respectively; the variance peak threshold is the product of the larger of the two intra-class variances and a preset coefficient, and the texture similarity threshold is the weighted average of the half-peak widths of the two main peaks; the preset coefficient is determined by segmenting the validation set to maximize the xylem intersection-union ratio, thereby avoiding the inconsistency of dimensions caused by directly using the grayscale valley value as the variance threshold; this invention solves the specific problem that fixed thresholds are prone to inaccurate segmentation when there are batch differences in the diameter of medicinal materials and lighting conditions, and improves the stability of the separation results of xylem and cortex regions under different working conditions.

[0055] Furthermore, the step of obtaining the xylem mask also includes: for the connected dark area determined to have hole features, extracting radial edge texture data within a preset width range outside its boundary, and detecting whether there are linear cracks extending outward from the hole boundary with a length exceeding a preset length threshold; when the linear crack is detected, marking the local area of ​​the corresponding hole boundary as a hole-associated tear area, allowing it to continue to participate in the subsequent extraction of the cut surface defect contour; when the linear crack is not detected, retaining the original hole feature determination of the connected dark area; Specifically, starting from the boundary of the dark area where the hole characteristics have been determined, the area is expanded outward by 1% to 3% of the equivalent diameter of the medicinal material end face to form a hole edge detection area. Edge detection is performed within this area, and linear textures extending from the hole boundary to the outer xylem are obtained using Hough line detection. Based on the statistical results of the length of the texture around normal holes, the 95th percentile is determined as a preset length threshold. Linear edges with a length exceeding the preset length threshold and extending from the hole boundary to the outer xylem are retained and included as part of the cut surface defect contour. This reduces the likelihood of natural hole edges being filtered out along with the hole after secondary tearing caused by cutting forces, and improves the completeness of defect contour extraction when natural holes and cutting defects coexist.

[0056] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A process control method based on multi-source sensor fusion, characterized in that, The method is applied to edge visual image early warning during the slicing process of root and rhizome medicinal materials, including: The surface texture image of the medicinal material before cutting and the processed end face image after cutting are obtained. The processed end face image is subjected to blur restoration processing, and the restored end face image with preserved contour details is output. The texture gradient of the longitudinal fibers is extracted from the surface texture image to generate the fiber direction vector; the probability of tearing along the grain is evaluated based on the spatial angle between the fiber direction vector and the preset cutting plane, and then converted into prior physical features. Region growing is performed on the restored end face image to separate the outer cortex region from the eccentrically grown xylem region; pore features are identified and filtered out within the xylem region to obtain a xylem mask; Sparse sampling is performed within the xylem mask area to locate the profile of the cut surface defect caused by cutting, and the edge fractal dimension features of the cut surface defect profile are extracted. The fiber orientation vector, prior physical features, and edge fractal dimension features are concatenated into a multi-source feature set and input into a pre-trained morphology classification network to generate and output corresponding process anomaly warning signals.

2. The processing whole-process control method based on multi-source sensor fusion according to claim 1, characterized in that: The steps for acquiring the processed end face image include: using a front-end vision acquisition device set in the feeding area to optically photograph the uncut surface of the medicinal material to acquire a surface texture image containing surface longitudinal wrinkle data; and using a rear-end vision acquisition device set behind the cutting station to capture the processed end face image after cutting.

3. The processing whole-process control method based on multi-source sensor fusion according to claim 2, characterized in that: The process of blur restoration includes: extracting edge pixels from the processing end face image, calculating the gray-level gradient change rate, extracting the spatial mapping direction with the largest gray-level gradient change rate as the edge blur direction feature; extracting the trailing diffusion length of the edge pixels as the blur scale feature; numerically reconstructing the edge blur direction feature and the blur scale feature to eliminate the dynamic trailing data in the processing end face image, and using it as the restored end face image.

4. The processing whole-process control method based on multi-source sensor fusion according to claim 3, characterized in that: The steps for obtaining the fiber orientation vector include: using the structure tensor operator to extract the orientation gradient data of grooves and cracks in the surface texture image at multiple scales; performing Gaussian smoothing filtering on the orientation gradient data and aggregating it to obtain the fiber orientation vector that characterizes the overall physical structure orientation of the medicinal material.

5. The processing whole-process control method based on multi-source sensor fusion according to claim 4, characterized in that: The steps for obtaining prior physical features include: calculating the spatial angle data between the fiber direction vector and the preset cutting plane normal vector in the established image space coordinate system; inputting the spatial angle data into a preset probability mapping function to obtain the longitudinal tear probability; comparing the longitudinal tear probability with a preset multi-level risk threshold range to obtain the discretized risk level corresponding to the longitudinal tear probability; using a one-hot encoding algorithm to convert the discretized risk level into a multi-dimensional vector format, and performing feature upscaling mapping calculation to generate a feature tensor with the same data dimension as the fiber direction vector, which serves as the prior physical feature.

6. The processing whole-process control method based on multi-source sensor fusion according to claim 1, characterized in that: The steps for separating the cortex and xylem regions include: calculating the local grayscale variance data of pixels in the acquired restored end-face image using a sliding window of a preset size; selecting multiple pixel nodes whose local grayscale variance data is greater than the peak variance threshold and converging them into an initial seed point set to adapt to the multi-core biological development characteristics of the eccentric growth of medicinal materials; starting from each pixel node in the initial seed point set, performing adaptive region growth operations on the outer pixels according to a preset texture similarity threshold, and simultaneously updating the region attribution label data of adjacent pixels; until the local regions grown from different initial seed points merge morphologically and stop the growth operation when they touch the grayscale gradient closed boundary, obtaining the fused growth region data; marking the fused growth region data as the eccentrically grown xylem region; and marking the range data outside the closed boundary up to the edge of the restored end-face image as the outer cortex region, thus completing the separation.

7. The processing whole-process control method based on multi-source sensor fusion according to claim 1, characterized in that: The steps for obtaining the xylem mask include: performing a global spatial search of low-grayscale connected dark area data within the marked xylem region to obtain an initial set of dark areas; when the grayscale mean of each connected dark area in the initial set of dark areas meets the environmental background threshold condition, tracking and calculating the boundary extension trajectory data of the connected dark areas; determining whether the boundary extension trajectory data is completely enclosed within the pixel coordinates of the xylem region and does not have any spatial intersection with the surrounding cortex region; if the determination result is yes, then the connected dark areas are determined to be pore features formed by natural physiological development inside the medicinal material; generating an initial binarization matrix with the same size as the restored end face image, filtering out the matrix elements corresponding to the pore features, and outputting the processed initial binarization matrix as the xylem mask.

8. The processing whole-process control method based on multi-source sensor fusion according to claim 1, characterized in that: The steps for extracting the edge fractal dimension features of the cut surface defect contour include: generating a sparse sampling grid matrix within the effective pixel coordinate range constrained by the xylem mask according to a preset sampling step size; extracting only the pixel grayscale data located at the intersection nodes of the sparse sampling grid matrix, and calculating macroscopic texture statistics data including grayscale skewness and kurtosis parameters. By comparing macroscopic texture statistics with preset end-face standard data, boundary regions with discontinuous gray-level abrupt changes are located, and the set of connected edge pixels is extracted as the cut surface defect contour. The box-counting dimension algorithm is used to analyze the topological features of the cut surface defect contour. Multiple square grids with continuously decreasing side lengths are used to cover the cut surface defect contour, and the number of grids required for complete coverage each time is counted. In the logarithmic coordinate system, the linear slope value between the reciprocal of the grid side length and the number of grids is fitted. The fitted linear slope value is used as the edge fractal dimension feature characterizing the degree of edge roughness and fracture.

9. The processing whole-process control method based on multi-source sensor fusion according to claim 8, characterized in that: The steps for generating the corresponding process anomaly early warning signal include: performing extreme value normalization on the extracted fiber direction vector and edge fractal dimension features respectively; combining the normalized fiber direction vector and edge fractal dimension features with prior physical features, and splicing and recombining them according to the preset tensor channel dimension to construct a multi-source feature tensor with unified dimension; inputting the multi-source feature tensor into a morphology classification network for nonlinear mapping and network weight calculation; outputting the probability distribution value of the current image state belonging to each morphology classification; selecting the category with the highest probability distribution value as the final morphology classification result, and generating a process anomaly early warning signal bound to the morphology classification result for terminal output.

10. The processing whole-process control method based on multi-source sensor fusion according to claim 9, characterized in that: The image state belongs to the following morphological classifications: excessive moisture loss, tearing under stress during cutting along the grain, and dulling of the cutting edge.

Images