Intelligent identification and sorting method and system for traditional Chinese medicinal materials

Abstract

The present application relates to the field of computer vision, in particular to a traditional Chinese medicinal material intelligent identification and sorting method and system. Based on a multispectral and microscopic vision fusion engine, sub-pixel level feature extraction and cross-modal knowledge distillation are used to achieve millisecond-level authenticity identification of traditional Chinese medicinal materials. PointRend point cloud segmentation and pharmacopoeia atlas adversarial verification are combined to completely eliminate similar medicinal material confusion. An innovative rotating YOLO-MoE multi-angle defect detection mechanism is used to achieve a zero confusion product control closed loop with zero manual intervention under the assistance of an adaptive guided execution unit, and the atlas database is continuously upgraded through a dynamic federated learning mechanism. The present application breaks through the technical bottleneck of existing traditional Chinese medicine recognition systems, which rely on manual experience and have weak authenticity identification capability, and completely solves key problems such as similar morphological medicinal material confusion caused by single visual recognition dimension in semi-automatic pharmacy and defect product missed detection caused by the lack of quality control mechanism. The present application improves the medicinal material identification and sorting process to a stable and reliable, self-adaptive and intelligent level.

Description

Technical Field

[0001] This invention relates to the field of computer vision, and in particular to a method and system for intelligent identification and sorting of Chinese medicinal materials. Background Technology

[0002] Traditional Chinese medicine pharmacies typically stock hundreds of medicinal herbs. Doctors' prescriptions often require multiple herbs, and traditional manual dispensing necessitates pharmacists frequently moving between cabinets to locate and weigh herbs, taking several minutes per prescription and prone to errors. While existing semi-automated pharmacies have introduced robotic arms and basic vision systems, they are still largely manual. The systems cannot identify variations in the morphology of bulk herbs, such as the confusion rate between white and red peony root, or counterfeit herbs, such as counterfeit astragalus root mixed in. Furthermore, they rely on manual replenishment and verification. Their vision modules, which only use RGB cameras to scan barcodes, lack sufficient accuracy in segmenting stacked and adhered herbs and have no quality control capabilities. For example, insect infestation and mold rely on visual inspection and secondary manual verification, resulting in low efficiency and a high risk of disputes. Summary of the Invention

[0003] This invention provides a method and system for intelligent identification and sorting of Chinese medicinal materials. The technical problem to be solved is that this invention effectively breaks through the technical bottleneck of existing Chinese medicinal material identification systems that rely on human experience and have weak ability to distinguish between genuine and counterfeit products. It completely solves the key problems of confusion of similar-shaped medicinal materials caused by the single visual recognition dimension in semi-automated pharmacies, missed detection of defective products caused by the lack of quality control mechanisms, and damage and residue of medicinal materials caused by the rigid execution of mechanical systems. At the same time, it bridges the gap between detection and execution links in the traditional sorting process, realizes the closed-loop processing of the entire chain from image perception to intelligent decision-making, and improves the identification and sorting process of medicinal materials to a stable, reliable, adaptive and collaborative intelligent level.

[0004] To achieve the above-mentioned objectives, this invention provides an intelligent identification and sorting method for Chinese medicinal materials, characterized by the following steps: S1: acquiring hyperspectral image data and visible light image data of the medicinal materials, performing baseline calibration processing on the hyperspectral image data to generate calibrated hyperspectral data; S2: inputting the calibrated hyperspectral data and visible light image data into a multi-scale feature alignment module to output spatially registered multimodal image data; S3: performing PointRend instance segmentation processing on the multimodal image data to obtain the three-dimensional point cloud coordinates and surface texture data of individual medicinal materials; S4: performing microstructure dynamic enhancement on the surface texture data to extract sub-pixel-level vessel pit feature vectors; S5: By using a cross-modal knowledge distillation model, the duct pit feature vector and hyperspectral molecular fingerprint feature are fused to generate a fused feature vector; S6: The fused feature vector is matched adversarially with the pharmacopoeia visual knowledge graph to output the medicinal material ID and authenticity confidence; S7: Based on the three-dimensional point cloud coordinates, a mechanical claw path planning instruction is generated, and the mechanical claw is driven to grab the medicinal material to the inspection table; S8: Rotational multi-angle defect scanning is performed on the inspection table, and the defect confidence is output through a transfer-enhanced YOLOv7-MoE model; S9: When the defect confidence is greater than a threshold, a sorting instruction is triggered and the abnormal sample library of the pharmacopoeia visual knowledge graph is updated. In this scheme, the threshold is set to 0.95.

[0005] Furthermore, the baseline calibration process in step S1 includes: acquiring dark current noise data, performing Savitzky-Golay filtering on the original hyperspectral data, and decomposing the spectral curve through non-negative matrix decomposition to remove ambient light interference components.

[0006] Furthermore, in step S4, a subpixel super-resolution convolutional network is used to upsample the surface texture data by 16 times; the edges of the duct pits are located through deformable convolutional layers to generate a binary pit mask.

[0007] Furthermore, the cross-modal knowledge distillation model in step S5 includes:

[0008] The teacher model uses a 3DResNeXt network with hyperspectral molecular fingerprint features as input; the student model uses a GraphAttention network with duct pit features as input; the output feature distributions of the teacher and student models are aligned using KL divergence loss.

[0009] Furthermore, the adversarial similarity matching in step S6 includes: constructing an adversarial generative network, whereby the generator synthesizes confusing medicinal material features, and the discriminator performs true / false classification; when the discriminator's misclassification rate for the synthesized features is <5%, training is stopped, and the discriminator weights are fixed for similarity matching.

[0010] Furthermore, in step S8, the rotation angle of the medicinal material on the detection stage is controlled to include 0°, 45°, 90°, and 135°; the multi-angle scan images are input into the transfer-enhanced YOLOv7-MoE model, wherein the transfer enhancement is to fine-tune the insect-boring microscopic dataset on the pre-trained weights.

[0011] Furthermore, the abnormal sample library update in step S9 includes: performing high-resolution tomographic scanning on the sorted defective medicinal materials to reconstruct a three-dimensional insect-infested channel model; and storing the insect-infested channel model associated with the medicinal material ID in the abnormal sample library of the pharmacopoeia visual knowledge graph.

[0012] Furthermore, when constructing the pharmacopoeia visual knowledge graph, federated learning is used to aggregate private data from multiple top-tier hospitals, and the graph parameters are updated in an encrypted manner.

[0013] This invention discloses an intelligent identification and sorting system for traditional Chinese medicinal materials, characterized by comprising: a spectral calibration unit for performing baseline calibration processing of hyperspectral data; a multimodal alignment unit for fusing calibrated hyperspectral data with visible light image data; a point cloud segmentation unit for generating three-dimensional point cloud coordinates and surface texture data of individual medicinal materials; a microscopic enhancement unit for extracting sub-pixel-level vessel and pit feature vectors; a knowledge distillation unit for fusing vessel and pit features with molecular fingerprint features; an adversarial matching unit for matching the true / false confidence scores of the pharmacopoeia visual knowledge graph; a path planning unit for generating the motion path of the mechanical gripper; a rotation scanning unit for driving the medicinal materials to rotate at multiple angles on the detection stage and performing scanning; a defect detection unit that integrates a transfer-enhanced YOLOv7-MoE model to output defect confidence scores; and a knowledge update unit for updating the abnormal sample database using Micro-CT data.

[0014] Furthermore, the micro-enhancement unit integrates a subpixel super-resolution convolutional network and a deformable convolutional layer; the knowledge update unit deploys a federated learning client to encrypt and upload the wormhole channel model to the cloud knowledge graph.

[0015] The beneficial effects of this invention are as follows: This invention, through multimodal visual fusion and intelligent decision-making mechanisms, completely solves the core problems in traditional Chinese medicine identification methods, such as morphological confusion, unclear identification of authenticity, and missed detection of defects, significantly improving the accuracy and efficiency of medicine sorting; its constructed closed-loop processing flow realizes automatic coordination from data acquisition to execution control, greatly reducing the need for manual intervention and operational complexity, while the dynamically updated pharmacopoeia knowledge graph continuously optimizes the system's identification capabilities, providing the Chinese medicine industry with a highly adaptable and reliable intelligent solution. Attached Figure Description

[0016] Figure 1 This is a system workflow diagram of the present invention;

[0017] Figure 2 Here is a system block diagram of the present invention: Figure 3 This is a system workflow diagram of Embodiment 1 of the present invention;

[0018] Figure 4 This is a system workflow diagram of Embodiment 2 of the present invention. Detailed Implementation

[0019] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. Identical components are indicated by the same reference numerals.

[0020] It should be noted that the terms “front,” “back,” “left,” “right,” “up,” and “down” used in the following description refer to the directions shown in the attached diagram, while the terms “inside” and “outside” refer to the directions toward or away from the geometric center of a specific component, respectively.

[0021] To make the content of this invention easier to understand, the technical solutions in the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings.

[0022] Example 1: Identification of genuine and counterfeit Angelica sinensis and detection of insect infestation.

[0023] In the scenarios of identifying genuine and counterfeit Angelica sinensis and detecting insect damage, firstly, the curled shape of Angelica sinensis leads to incomplete 3D point cloud reconstruction, and conventional segmentation algorithms cannot accurately capture the boundaries; furthermore, counterfeit Angelica pubescens has high similarity to Angelica sinensis, with similar spectral characteristics and subtle differences in microscopic vessel pits, making it easy to confuse with traditional single-modal identification; then, the size of the borer holes in the early stage of insect damage is small and mostly hidden in the internal vessel network, making it difficult to effectively identify early damage characteristics using conventional visible light scanning and CT slice thickness, while the irregular topological structure of the borer tunnels increases the complexity of 3D modeling and quantitative analysis.

[0024] like Figure 1 The specific steps include:

[0025] S1. Simultaneously acquire raw image data of Angelica sinensis samples using a hyperspectral camera and a visible light camera; then, perform baseline calibration on the hyperspectral data: acquire dark current noise data as a benchmark, perform Savitzky-Golay filtering on the raw hyperspectral data to smooth the noise; next, decompose the spectral curve using a non-negative matrix factorization algorithm to remove ambient light interference components, and output calibrated hyperspectral data to ensure data quality and lay the foundation for subsequent multimodal fusion;

[0026] In this embodiment, the wavelength range of the hyperspectral camera is 900-1700nm; the Savitzky-Golay filtering algorithm is configured with a window size of 15 points and a third-order fitting, that is, when calculating a point, the data of 15 adjacent wavelength points before and after it are considered, and a cubic polynomial is fitted on these points for smoothing.

[0027] S2. The calibrated hyperspectral data and visible light image data are input into a multi-scale feature alignment module. This module performs processing through a pyramid feature alignment network: extracting multi-scale features from the visible light image and simultaneously extracting the spatial-spectral joint features from the hyperspectral data; then, performing adaptive cross-correlation alignment at multiple scale levels, using a Gaussian weighted kernel to adjust feature position deviations, and finally outputting spatially registered multimodal image data. This achieves accurate alignment of different modal data and solves the feature misalignment problem caused by curled morphology.

[0028] S3. Perform PointRend instance segmentation on the registered multimodal image data. This process employs an iterative point sampling mechanism: first, an initial mask is generated through coarse segmentation; then, refined sampling is performed on the difficult point set at the boundary, outputting the 3D point cloud coordinates and surface texture data of the individual medicinal materials. The segmentation loss function combines mask loss and point sampling loss to ensure the complete reconstruction of the curled morphology, providing precise positional information for the robotic gripper.

[0029] S4. Dynamic enhancement of surface texture data to reveal microstructure. First, a subpixel super-resolution convolutional network is used to upsample the texture data by 16 times to generate a high-resolution image. By magnifying microscopic details, it is easier to distinguish the subtle differences between Angelica sinensis and counterfeit products. Then, deformable convolutional layers are used to locate the edges of vessel pits, generate a binary pit mask, and extract subpixel-level vessel pit feature vectors.

[0030] S5. By fusing duct pit features and hyperspectral molecular fingerprint features through a cross-modal knowledge distillation model, multi-source information is integrated to enhance the ability to distinguish between genuine and fake products. In this embodiment, the teacher model is a 3DResNeXt network used to process hyperspectral molecular fingerprint features, and the student model is a GraphAttention network used to process duct pit features. KL divergence loss is used to align the output feature distributions of the two models. In this embodiment, the temperature parameter is set to 3, and finally, a fused feature vector is generated.

[0031] S6. Perform adversarial similarity matching between the fused feature vector and the pharmacopoeia visual knowledge graph to solve the high confusion problem between Angelica sinensis and Angelica pubescens; construct an adversarial generative network, where the generator synthesizes features of counterfeit Angelica pubescens and the discriminator performs true and false classification; during training, stop training when the discriminator's misclassification rate of the synthesized features is less than 5%, fix the discriminator weights to perform similarity matching, and output the herb ID and the confidence level of authenticity.

[0032] S7. Generate path planning instructions for the robotic gripper based on 3D point cloud coordinates. Calculate the centroid coordinates of the medicinal material and ensure safety through collision avoidance detection. In this embodiment, the maximum circumscribed sphere radius is calculated, maintaining a safe distance of 5mm. Select the point with the most consistent normal vector as the gripping point, generate a joint angle sequence, and drive the robotic gripper to grip the medicinal material onto the inspection table, achieving automatic gripping and avoiding manual intervention.

[0033] S8. Perform a rotating multi-angle defect scan on the inspection platform. Control the medicinal material to rotate at multiple angles on the inspection platform, including 0°, 45°, 90°, and 135°, and pause at each angle for 0.5 seconds for scanning to cover multi-angle insect infestation detection and avoid missed detections; input the multi-angle scan images into the transfer-enhanced YOLOv7-MoE model, fine-tune the insect infestation microscopic dataset, and output the defect confidence score through dynamic routing via an expert network.

[0034] S9. When the defect confidence level is greater than the threshold of 0.95, a sorting instruction is triggered. High-resolution tomographic scanning is performed on the sorted defective medicinal materials to reconstruct a three-dimensional insect-infested channel model. This model is associated with the medicinal material ID and stored in the abnormal sample database of the pharmacopoeia visual knowledge graph. The graph parameters are then encrypted and updated using a federated learning mechanism. These graph parameters can aggregate data from multiple hospitals.

[0035] This embodiment solves the problems of distorted segmentation of Angelica sinensis with curled morphology, high confusion with counterfeit Angelica pubescens, and difficulty in detecting insect-damaged defects. By combining hyperspectral molecular fingerprinting with sub-pixel-level pit features through cross-modal distillation fusion, the accuracy of Angelica sinensis authenticity identification is improved. Experiments show that the accuracy rate of Angelica sinensis authenticity identification reaches 99%, which is 15% higher than traditional detection methods, essentially eliminating the confusion with Angelica pubescens. Through rotational multi-angle scanning combined with the transfer-enhanced YOLOv7-MoE model, micron-level insect-damaged channels are detected. Experiments show that the false negative rate for insect-damaged channels is <0.5%, and the single-sample detection time is controlled within 1 second. Through 3D point cloud segmentation and mechanical claw path planning, sorting without human intervention can be achieved. Based on high-resolution tomographic reconstruction of the insect-damaged model, the knowledge graph is dynamically updated through federated learning, reducing the system's false positive rate. Experiments show that the false positive rate is reduced by more than 20%.

[0036] Example 2: Rapid differentiation between white peony root and red peony root

[0037] In the sorting of Chinese herbal medicine slices, there is a high probability of morphological confusion between white peony and red peony, as they both belong to the genus Paeonia in the Ranunculaceae family. Both slices are round or oval, and their color after drying falls into the gray-white to brownish-red spectrum. Traditional manual sorting relies on the difference in lenticel distribution, with white peony having fewer and lighter lenticels and red peony having more and darker lenticels. However, because the RGB vision system cannot penetrate the surface to analyze the internal microstructure, such as the cork layer stone cell rings unique to red peony and the high-density calcium oxalate clusters of white peony, the confusion rate is relatively high in actual sorting. Both share the characteristic absorption peak of paeoniflorin, while the content of paeonol unique to red peony is too low to produce an effective spectral difference, resulting in a significant decrease in the accuracy of traditional near-infrared identification. Furthermore, the common insect infestation defects in the pith of red peony are easily confused with the normal lenticels of white peony in two-dimensional images, further amplifying the sorting error.

[0038] like Figure 2 The specific steps include:

[0039] S1. Similar to Example 1, raw image data of white peony and red peony samples are acquired simultaneously using a hyperspectral camera and a visible light camera; then, the hyperspectral data is baseline-calibrated, and then ambient light interference components are removed by a non-negative matrix to output the calibrated hyperspectral data.

[0040] S2, similar to Example 1, involves inputting calibrated hyperspectral data and visible light image data into a pyramid feature alignment network. Then, adaptive cross-correlation alignment is performed at multiple scale levels, using Gaussian weighted kernels to adjust feature position deviations, ultimately outputting spatially registered multimodal image data. This achieves accurate alignment of different modalities, resolving feature misalignment issues caused by curled morphology.

[0041] S3. Similar to the embodiment, PointRend instance segmentation is used to process multimodal image data, and iterative refinement is performed on difficult sampling boundary points, including the edge of peony slices, to output three-dimensional point cloud coordinates and surface texture data.

[0042] S4. Dynamic enhancement of surface texture data using microstructure. First, a subpixel super-resolution convolutional network is used to upsample the texture data by 16 times to generate a high-resolution image; by magnifying microscopic details, it is easier to distinguish the subtle differences between Angelica sinensis and its adulterants; unlike Example 1, for Paeonia lactiflora, deformable convolutional layers are used to locate the cork layer stone cell rings, and for Paeonia veitchii, the distribution characteristics of calcium oxalate cluster crystals are extracted.

[0043] S5. By fusing duct pit features and hyperspectral molecular fingerprint features through a cross-modal knowledge distillation model, multi-source information is integrated to enhance the ability to distinguish between genuine and counterfeit products. In this embodiment, the teacher model is a 3DResNeXt network used to focus on the paeoniflorin feature peak, and the student model is a GraphAttention network used to enhance the paeonol texture identification. The output feature distributions of the two models are aligned using KL divergence loss. In this embodiment, the temperature parameter is set to 3, and finally, a fused feature vector is generated.

[0044] S6. Perform adversarial similarity matching between the fused feature vector and the pharmacopoeia visual knowledge graph; construct an adversarial generative network, and the generator synthesizes confusing features. In this embodiment, the white peony spectrum and red peony texture are mixed; during training, the discriminator strengthens the matching of key identifiers. In this embodiment, the key discriminant features of white peony are: vascular diameter > 40 μm, cluster crystal density < 5 / mm²; the key discriminant features of red peony are: cork layer stone cell rings; when the discriminator's misclassification rate for the synthesized features is less than 5%, stop training and output the medicinal material ID and the confidence level of authenticity.

[0045] S7. Generate path planning instructions for the robotic gripper based on 3D point cloud coordinates. Calculate the centroid coordinates of the medicinal material and ensure safety through collision avoidance detection. In this embodiment, the maximum circumscribed sphere radius is calculated, maintaining a safe distance of 5mm. Select the point with the most consistent normal vector as the gripping point, generate a joint angle sequence, and drive the robotic gripper to grip the medicinal material onto the inspection table, achieving automatic gripping and avoiding manual intervention.

[0046] S8. Perform a rotating multi-angle defect scan on the inspection table. Control the medicinal material to rotate at multiple angles on the inspection table, including 0° and 90°, and pause at each angle for 0.3 seconds for scanning;

[0047] Unlike Example 1, this method optimizes the use of orthogonal dual-angle scanning at 0° and 90°. The orthogonal dual angles can cover the entire planar area. The surface texture is viewed from the front at 0° and the edge of the slice is viewed from the side at 90°. This avoids the problem of needing to fill in the blanks of the blocky medicinal materials due to the curved surface. Compared with the blocky medicinal materials of Angelica sinensis in Example 1, the redundant angles are reduced by 45° and 135°.

[0048] The multi-angle scanned images are input into the transfer-enhanced YOLOv7-MoE model. In this embodiment, the images also include a pre-trained weight fine-tuning dataset of 10,000 peony defect images, which are dynamically routed to an expert network for insect infestation and mold, and the defect confidence is output.

[0049] S9. When the defect confidence level is greater than the threshold of 0.95, a sorting instruction is triggered. High-resolution tomographic scanning is performed on the sorted defective medicinal materials to reconstruct a three-dimensional distribution model of stone cells. This model is associated with the medicinal material ID and stored in the pharmacopoeia visual knowledge graph. A topological map of peony stone cells is also associated with the medicinal material ID and stored. The map parameters can aggregate data from multiple hospitals.

[0050] This embodiment achieves millisecond-level precise differentiation between white peony and red peony through multimodal fusion of hyperspectral and microscopic texture combined with an adversarial similarity matching mechanism. Targeting the high morphological and spectral similarity between the two, the microscopic enhancement unit specifically extracts the cork cell rings of red peony and the calcium oxalate cluster distribution of white peony. After cross-modal distillation and fusion, these are input into an adversarial matching network. The generator synthesizes confusion features to induce the discriminator to strengthen key identifier learning, thereby improving the identification accuracy. Experiments show that the accuracy is improved by over 99%, and the identification time for a single sample is controlled within 1 second. Simultaneously executed defect detection employs dual-angle optimized scanning, combined with the Paeonia suffruticosa-specific YOLOv7-MoE model, and integrates precise grasping by a robotic gripper and federated knowledge updates to construct a fully closed-loop system covering variety identification, defect screening, and knowledge evolution. This significantly reduces the misjudgment rate and resolves the risk of confusion caused by defect interference in traditional sorting technologies.

[0051] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent identification and sorting of Chinese medicinal materials, characterized in that, Includes the following steps: S1: Acquire hyperspectral image data and visible light image data of medicinal materials, perform baseline calibration processing on the hyperspectral image data, and generate calibrated hyperspectral data; S2: Input the calibrated hyperspectral data and visible light image data into the multi-scale feature alignment module, and output spatially registered multimodal image data; S3: Perform PointRend instance segmentation processing on the multimodal image data, generate an initial mask using an iterative point sampling mechanism, refine the sampling for difficult boundary point sets, and obtain the three-dimensional point cloud coordinates and surface texture data of the independent medicinal materials. S4: Perform microstructure dynamic enhancement on the surface texture data, use a subpixel super-resolution convolutional network for upsampling, locate the edges of the duct pits through deformable convolutional layers and generate a binary pit mask, thereby extracting subpixel level duct pit feature vectors. S5: By using a cross-modal knowledge distillation model, the duct pit feature vector and the hyperspectral molecular fingerprint feature are fused to generate a fused feature vector; in the cross-modal knowledge distillation model, the 3DResNeXt network for processing hyperspectral molecular fingerprint features is used as the teacher model, and the Graph Attention network for processing duct pit features is used as the student model, and the output feature distributions of the teacher model and the student model are aligned by KL divergence loss. S6: Perform adversarial similarity matching between the fused feature vector and the pharmacopoeia visual knowledge graph, and output the medicinal material ID and the confidence level of authenticity; S7: Generate a mechanical claw path planning instruction based on the three-dimensional point cloud coordinates, and drive the mechanical claw to grab the medicinal material to the testing table; S8: Perform a rotating multi-angle defect scan on the detection stage and output the defect confidence level through the transfer-enhanced YOLOv7-MoE model; S9: When the confidence level of the defect is greater than the threshold, a sorting instruction is triggered to perform a high-resolution tomographic scan on the defective medicinal material to reconstruct a three-dimensional model. The three-dimensional model is then associated with the medicinal material ID and stored in the abnormal sample library of the pharmacopoeia visual knowledge graph. The graph parameters are dynamically updated through a federated learning mechanism to form a closed-loop processing flow that covers identification and knowledge evolution.

2. The intelligent identification and sorting method for Chinese medicinal materials according to claim 1, characterized in that, The baseline calibration process in step S1 includes: Dark current noise data were collected, and Savitzky-Golay filtering was performed on the raw hyperspectral data. The spectral curve is decomposed by nonnegative matrix decomposition to eliminate ambient light interference components.

3. The intelligent identification and sorting method for Chinese medicinal materials according to claim 1, characterized in that, The adversarial similarity matching in step S6 includes: A generative adversarial network is constructed, in which the generator synthesizes confusing medicinal features, and the discriminator performs true and false classification. When the misclassification rate of the discriminator for the synthesized features is less than a threshold, training is stopped, and the discriminator weights are fixed for similarity matching.

4. The intelligent identification and sorting method for Chinese medicinal materials according to claim 1, characterized in that, Step S8 includes: controlling the medicinal material to rotate at multiple angles on the testing platform; Multi-angle scanned images are input into a transfer-enhanced YOLOv7-MoE model, whereby the transfer enhancement involves fine-tuning the insect-infested microscopic dataset on pre-trained weights.

5. A smart identification and sorting system for traditional Chinese medicinal materials, characterized in that, include: The spectral calibration unit is used to perform baseline calibration processing of hyperspectral data; A multimodal alignment unit is used to fuse calibrated hyperspectral data with visible light image data; Point cloud segmentation unit, used to generate three-dimensional point cloud coordinates and surface texture data of individual medicinal materials; Microscopic enhancement unit, used to extract sub-pixel level vessel pit feature vectors; Knowledge distillation unit, used to fuse duct pit features with molecular fingerprint features; Adversarial matching unit is used to match the confidence scores of the pharmacopoeia visual knowledge graph output; The path planning unit is used to generate the motion path of the robotic gripper. A rotary scanning unit is used to drive the medicinal materials to rotate at multiple angles on the testing stage; The defect detection unit integrates a transfer-enhanced YOLOv7-MoE model to output defect confidence scores. The knowledge update unit is used to update the abnormal sample library using Micro-CT data.

6. The system according to claim 5, characterized in that: The micro-enhancing unit integrates a subpixel super-resolution convolutional network and a deformable convolutional layer; The knowledge update unit deploys a federated learning client to encrypt and upload the wormhole channel model to the cloud knowledge graph.

Images