Veterinary traditional Chinese medicine prescription microscopic image identification system based on deep learning

Abstract

The application discloses to the field of veterinary traditional Chinese medicine identification technology and deep learning cross technology, and particularly relates to a veterinary traditional Chinese medicine prescription microscopic image identification system based on deep learning, which first extracts shallow microscopic features in the image based on an effective area microscopic image set, then adopts an improved convolutional neural network model to extract deep semantic features, and finally adopts an attention mechanism fusion algorithm to adaptively fuse the shallow microscopic features and the deep semantic features; based on this, the application can accurately adapt to veterinary traditional Chinese medicine prescription microscopic images, can comprehensively extract fine features and component correlation features such as cell morphology, tissue structure and inclusions in the image, can effectively avoid multi-component feature superposition interference, and can adapt to different types and different compatibility of veterinary traditional Chinese medicine prescriptions.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of identification of veterinary Chinese medicine and deep learning, specifically a microscopic image identification system for veterinary Chinese medicine prescriptions based on deep learning. Background Technology

[0002] Veterinary traditional Chinese medicine formulas, with their advantages of fewer side effects, lower risk of drug resistance, and compatibility with animal physiological characteristics, are widely used in the prevention, treatment, and health maintenance of livestock and poultry diseases. Their quality directly affects livestock farming efficiency, animal product safety, and public health safety. Microscopic identification, as one of the most core and fundamental techniques for identifying veterinary traditional Chinese medicine formulas, is based on observing the microscopic characteristics of each herbal component (such as cell morphology, tissue structure, and inclusions) and combining this with the component characteristics to identify the authenticity and grade the quality of the formula.

[0003] With the large-scale and standardized development of animal husbandry, the usage of traditional Chinese medicine prescriptions for veterinary use has increased significantly, leading to a growing demand for higher efficiency and accuracy in identification. Currently, deep learning technology, with its powerful feature extraction and pattern recognition capabilities, has achieved mature applications in fields such as the identification of medicinal herbs and the testing of agricultural products.

[0004] However, current microscopic image identification techniques for veterinary Chinese herbal medicine formulas may suffer from poor adaptability to existing image recognition technologies and incomplete feature extraction. This is because some existing microscopic image identification techniques for veterinary Chinese herbal medicine formulas use traditional image recognition algorithms (such as threshold segmentation and edge detection) without incorporating the advantages of deep learning technology. These techniques cannot effectively extract deep features from microscopic images and are not optimized for the compatibility characteristics of veterinary Chinese herbal medicine formulas, resulting in poor adaptability. For example, traditional image recognition techniques can only extract shallow contour features from microscopic images and cannot identify subtle features such as calcium oxalate crystals in licorice cells and starch granules in dried ginger cells in "Gancao Xiexin Tang" (Licorice Decoction for Clearing the Heart), leading to an inability to distinguish between licorice and adulterants (such as Astragalus). Furthermore, traditional algorithms do not consider the feature superposition problem after mixing multiple components in veterinary Chinese herbal medicine formulas. When a formula contains more than three Chinese herbal medicine components, feature interference is severe, and identification accuracy decreases. Summary of the Invention

[0005] To address the aforementioned technical problems of poor adaptability and incomplete feature extraction in existing image recognition technologies, this invention provides the following technical solution:

[0006] A deep learning-based microscopic image identification system for veterinary traditional Chinese medicine prescriptions includes:

[0007] The microscopic image acquisition and preprocessing module acquires microscopic images of veterinary Chinese medicine prescriptions and outputs a set of effective region microscopic images through noise removal, standardization, and segmentation operations.

[0008] The microscopic feature extraction module includes a shallow feature extraction unit, a deep feature extraction unit, and a feature fusion unit. First, the shallow feature extraction unit extracts shallow microscopic features from the effective region microscopic image set output by the microscopic image acquisition and preprocessing module, and outputs a shallow microscopic feature vector set. Next, the deep feature extraction unit, based on the shallow microscopic feature vector set output by the shallow feature extraction unit and combined with the effective region microscopic image set output by the microscopic image acquisition and preprocessing module, uses an improved convolutional neural network model to extract deep semantic features, and outputs a deep semantic feature vector set. Finally, the feature fusion unit, based on the shallow microscopic feature vector set output by the shallow feature extraction unit and the deep semantic feature vector set output by the deep feature extraction unit, uses an attention mechanism fusion algorithm to adaptively fuse the shallow microscopic features and deep semantic features; simultaneously, it performs dimensionality reduction processing on the fused feature vectors and outputs a fused feature vector set.

[0009] The veterinary Chinese medicine component identification module identifies each Chinese medicine component in the veterinary Chinese medicine formula based on the fusion feature vector set output by the microscopic feature extraction module, determines the type and purity of the components, and outputs a component identification result table.

[0010] The formula identification and quality assessment module, based on the component identification result table output by the veterinary Chinese medicine component identification module, combined with the compatibility standards of veterinary Chinese medicine formulas, realizes the identification of the authenticity of the formula and the assessment of its quality grade, and outputs the formula authenticity identification result table and the formula quality grade assessment result table.

[0011] The anomaly warning and traceability module, based on the formula authenticity identification result table and formula quality grade assessment result table output by the formula identification and quality assessment module, identifies abnormal situations of the formula and issues warnings, while realizing full-process traceability of the formula and outputting an anomaly warning information table and a full-process traceability report of the formula.

[0012] The model adaptive optimization and intelligent application module for identification results, based on the component identification result table output by the veterinary Chinese medicine component identification module and the formula authenticity identification result table and formula quality grade evaluation result table output by the formula identification and quality evaluation module, performs dynamic adaptive optimization of the deep learning identification model, and transforms the identification and quality evaluation results into a feasible intelligent application solution.

[0013] As a preferred embodiment of the deep learning-based microscopic image identification system for veterinary traditional Chinese medicine prescriptions described in this invention, the microscopic image acquisition and preprocessing module includes:

[0014] The microscopic image acquisition unit, based on the sample characteristics of veterinary Chinese medicine prescriptions, sets standardized acquisition parameters to acquire images of prescription samples from different batches and different parts in all directions, and outputs a set of original microscopic images.

[0015] The image noise removal unit identifies various types of noise in the image based on the original set of microscopic images output by the microscopic image acquisition unit. At the same time, it adopts an adaptive noise removal algorithm to set differentiated filtering parameters for different types of noise. While removing noise, it retains the subtle features in the microscopic image and outputs a set of denoising microscopic images.

[0016] The image normalization unit performs normalization processing on the set of denoised microscopic images output by the image noise removal unit; at the same time, it performs geometric correction on the images and outputs a set of normalized microscopic images.

[0017] The image segmentation unit, based on the standardized set of microscopic images output by the image standardization unit, uses a deep learning semantic segmentation algorithm to automatically segment the microscopic images, separating the effective discriminative regions from the background regions in the images, and outputting a set of effective region microscopic images.

[0018] As a preferred embodiment of the deep learning-based microscopic image identification system for veterinary traditional Chinese medicine prescriptions according to the present invention, the veterinary traditional Chinese medicine component identification module includes:

[0019] The component identification model training unit constructs a sample library of veterinary Chinese medicine components containing standard microscopic images and corresponding feature vectors. Based on the fusion feature vector set output by the microscopic feature extraction module, combined with the standard data in the sample library, the component identification model is trained, and the trained component identification model and the veterinary Chinese medicine component sample library are output.

[0020] The component preliminary identification unit, based on the component identification model trained by the component identification model training unit and the fused feature vector set output by the microscopic feature extraction module, inputs the fused feature vector into the component identification model to perform preliminary identification, determines the type of Chinese medicine component corresponding to each effective region image, calculates the identification confidence, marks the identification results of the confidence, and outputs a preliminary component identification result table and a set of low-confidence identification images.

[0021] The secondary component identification and verification unit, based on the preliminary component identification result table and low-confidence identification image set output by the preliminary component identification unit, and combined with the fusion feature vector set output by the microscopic feature extraction module and the veterinary Chinese medicine component sample library output by the component identification model training unit, performs secondary identification by feature comparison and manual verification; at the same time, it integrates the preliminary identification results and the secondary identification results and outputs the component identification result table.

[0022] As a preferred embodiment of the deep learning-based microscopic image identification system for veterinary traditional Chinese medicine prescriptions according to the present invention, the prescription identification and quality assessment module includes:

[0023] The formula compatibility standard library construction unit collects standard compatibility information of veterinary Chinese medicine formulas, and standardizes and organizes the compatibility information to construct a veterinary Chinese medicine formula compatibility standard library;

[0024] The formula authenticity identification unit, based on the veterinary Chinese medicine formula compatibility standard library constructed by the formula compatibility standard library construction unit and the component identification result table output by the veterinary Chinese medicine component identification module, performs formula authenticity identification and outputs formula authenticity identification result table.

[0025] The formula quality grade assessment unit, based on the formula authenticity identification result table output by the formula authenticity identification unit and the component identification result table output by the veterinary Chinese medicine component identification module, and combined with the veterinary Chinese medicine formula compatibility standard library constructed by the formula compatibility standard library construction unit, assesses the quality grade of qualified formulas and outputs a formula quality grade assessment result table.

[0026] As a preferred embodiment of the deep learning-based microscopic image identification system for veterinary traditional Chinese medicine prescriptions described in this invention, the anomaly warning and tracing module includes:

[0027] The anomaly warning unit, based on the formula authenticity identification result table and formula quality grade evaluation result table output by the formula identification and quality evaluation module, sets an anomaly warning threshold and monitors the identification and evaluation results in real time. When an anomaly exceeding the warning threshold is detected, it automatically issues a warning signal of the corresponding type and marks the warning level. At the same time, it analyzes the cause of the anomaly and outputs an anomaly warning information table.

[0028] The full-process traceability unit, based on the abnormal warning information table output by the abnormal warning unit, the set of effective area microscopic images output by the microscopic image acquisition and preprocessing module, and the formula authenticity identification result table output by the formula identification and quality assessment module, combined with the formula compatibility standard, performs full-process traceability of the formula and outputs a full-process traceability report and a traceability query record table.

[0029] As a preferred embodiment of the deep learning-based microscopic image identification system for veterinary traditional Chinese medicine prescriptions described in this invention, the model adaptive optimization and intelligent application module for identification results includes:

[0030] The real-time performance monitoring unit for the identification model monitors the model's performance in real time based on the component identification result table output by the veterinary Chinese medicine component identification module and the formula authenticity identification result table and formula quality grade evaluation result table output by the formula identification and quality assessment module. Simultaneously, it sets performance evaluation indicators, compares the current performance indicators with preset standard thresholds in real time, records performance fluctuation data, identifies scenarios of model performance decline, and outputs a real-time model performance monitoring report, performance anomaly warning signal, and preliminary analysis table of the causes of model performance fluctuations.

[0031] The model adaptive optimization unit, based on the real-time performance monitoring report, performance anomaly warning signal and preliminary analysis table of model performance fluctuation reasons output by the real-time performance monitoring unit of the identification model, combined with the fusion feature vector set and prescription compatibility standard output by the microscopic feature extraction module, performs adaptive optimization of the model and outputs the optimized model and the optimized model performance prediction report.

[0032] The intelligent transformation unit for identification results, based on the optimized model performance prediction report output by the model adaptive optimization unit, combined with the formula authenticity identification result table and formula quality grade evaluation result table output by the formula identification and quality assessment module, and the abnormal warning information table output by the abnormal warning and traceability module, transforms the identification results into a feasible intelligent application solution; at the same time, it standardizes and organizes the identification results and application solutions to form an exportable report template, and outputs a feasibility analysis table for the implementation of the application solution;

[0033] The optimization model verification and iteration unit, based on the optimized model output by the model adaptive optimization unit and the application scheme feasibility analysis table output by the intelligent conversion unit of the identification results, combined with the new batch of effective area microscopic images output by the microscopic image acquisition and preprocessing module, compares the verification results with the manual identification results, calculates the actual performance index of the optimized model, and determines whether it meets the preset standard; at the same time, it collects feedback data after the application scheme is implemented.

[0034] Compared with existing technologies:

[0035] 1. By employing a deep learning semantic segmentation algorithm and an improved convolutional neural network model, combined with an attention mechanism fusion algorithm to integrate shallow and deep microscopic features, this method achieves precise adaptation to microscopic images of veterinary Chinese herbal medicine formulas. It can comprehensively extract subtle features such as cell morphology, tissue structure, and contents, as well as the correlation features between components in the image, effectively avoiding interference from the superposition of multi-component features. It can adapt to different types and combinations of veterinary Chinese herbal medicine formulas, significantly improving the problem that traditional image recognition can only extract shallow features and has insufficient adaptability, thus enhancing the comprehensiveness and targeting of feature extraction.

[0036] 2. By setting unified microscopic image acquisition parameters, preprocessing standards and identification procedures, a standardized veterinary Chinese medicine prescription compatibility standard library and component feature library are constructed. The operation parameters and judgment standards of each link are standardized, which can realize the standardization of the whole process identification, eliminate the differences caused by different scenarios and different operations, improve the consistency and repeatability of identification results, and support unified identification applications across institutions and scenarios.

[0037] 3. By establishing the correlation between the characteristics and quality of each component of veterinary Chinese medicine prescriptions, integrating data from the entire process of image acquisition, feature extraction, identification and evaluation, and combining it with an anomaly threshold monitoring mechanism, it can realize component correlation analysis and quality traceability, quickly identify abnormal conditions of prescriptions and issue accurate early warnings, clarify the source of anomalies, and provide reliable support for quality control and responsibility traceability. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the overall framework of the present invention;

[0039] Figure 2 This is a schematic diagram of the framework of the microscopic image acquisition and preprocessing module of the present invention;

[0040] Figure 3 This is a schematic diagram of the microscopic feature extraction module framework of the present invention;

[0041] Figure 4 This is a schematic diagram of the framework of the veterinary Chinese medicine component identification module of the present invention;

[0042] Figure 5 This is a schematic diagram of the prescription identification and quality assessment module of the present invention;

[0043] Figure 6 This is a schematic diagram of the anomaly warning and tracing module framework of the present invention;

[0044] Figure 7 This is a schematic diagram of the intelligent application module framework for adaptive optimization and identification results of the model in this invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0046] This invention provides a deep learning-based microscopic image identification system for veterinary traditional Chinese medicine prescriptions. Please refer to [link / reference]. Figure 1 ,include:

[0047] The microscopic image acquisition and preprocessing module acquires microscopic images of veterinary Chinese medicine prescriptions and outputs a set of effective region microscopic images through noise removal, standardization, and segmentation operations.

[0048] The microscopic feature extraction module extracts the shallow microscopic features and deep semantic features of each Chinese herbal medicine component in the prescription based on the set of effective area microscopic images output by the microscopic image acquisition and preprocessing module, and outputs a set of fused feature vectors after feature fusion.

[0049] The veterinary Chinese medicine component identification module identifies each Chinese medicine component in the veterinary Chinese medicine formula based on the fusion feature vector set output by the microscopic feature extraction module, determines the type and purity of the components, and outputs a component identification result table.

[0050] The formula identification and quality assessment module, based on the component identification result table output by the veterinary Chinese medicine component identification module, combined with the compatibility standards of veterinary Chinese medicine formulas, realizes the identification of the authenticity of the formula and the assessment of its quality grade, and outputs the formula authenticity identification result table and the formula quality grade assessment result table.

[0051] The anomaly warning and traceability module, based on the formula authenticity identification result table and formula quality grade assessment result table output by the formula identification and quality assessment module, identifies abnormal situations of the formula and issues warnings, while realizing full-process traceability of the formula and outputting an anomaly warning information table and a full-process traceability report of the formula.

[0052] The model adaptive optimization and intelligent application module for identification results, based on the component identification result table output by the veterinary Chinese medicine component identification module and the formula authenticity identification result table and formula quality grade evaluation result table output by the formula identification and quality evaluation module, performs dynamic adaptive optimization of the deep learning identification model, and transforms the identification and quality evaluation results into a feasible intelligent application solution.

[0053] Please see Figure 2 The microscopic image acquisition and preprocessing module includes:

[0054] The microscopic image acquisition unit, based on the sample characteristics (powder, granules, etc.) of veterinary Chinese medicine prescriptions, sets standardized acquisition parameters (magnification, light intensity, focal length, exposure time) to enable comprehensive acquisition of prescription sample images from different batches and parts through the linkage of the microscope and image acquisition equipment. During the acquisition process, it automatically records basic sample information (sample name, acquisition time, batch number, acquisition personnel, sample source) to ensure the comprehensiveness and traceability of the acquired images and outputs a set of original microscopic images.

[0055] The image noise removal unit, based on the original set of microscopic images output by the microscopic image acquisition unit, identifies various types of noise in the images (such as Gaussian noise, salt-and-pepper noise, and noise caused by uneven illumination). At the same time, it adopts an adaptive noise removal algorithm (an improved algorithm combining median filtering and Gaussian filtering) to set differentiated filtering parameters for different types of noise. While removing noise, it preserves the subtle features in the microscopic images (such as intracellular contents and tissue texture) to avoid feature loss, and outputs a set of denoising microscopic images.

[0056] The image standardization unit performs standardization processing on the set of denoised microscopic images output by the image noise removal unit: unifying image resolution (adjusted to 1024×1024 pixels), unifying image grayscale value range (0-255), and unifying image contrast and brightness (based on an adaptive histogram equalization algorithm), eliminating image differences caused by different acquisition parameters and different sample states, and ensuring that all images are in the same standard dimension; at the same time, geometric correction is performed on the images to correct image distortion caused by microscope angle deviation and sample placement tilt during the acquisition process; and outputs a standardized set of microscopic images;

[0057] The image segmentation unit, based on the standardized set of microscopic images output by the image standardization unit, uses a deep learning semantic segmentation algorithm (improved U-Net algorithm) to automatically segment the microscopic images, separating the effective identification regions (such as Chinese medicine cells, tissues, and inclusions) from the background regions (such as slides, bubbles, and impurity shadows). During the segmentation process, the segmentation threshold is optimized by combining the characteristics of the microscopic features of veterinary Chinese medicine to ensure the integrity and accuracy of the effective regions, and outputs a set of microscopic images of the effective regions.

[0058] Please see Figure 3 The microscopic feature extraction module includes:

[0059] The shallow feature extraction unit extracts shallow microscopic features from the effective area microscopic image set output by the microscopic image acquisition and preprocessing module. These features include: contour features (perimeter, area, roundness, edge complexity), texture features (gray-level co-occurrence matrix, entropy value, contrast), and color features (gray-level mean, gray-level variance, color histogram). Simultaneously, it sets differentiated shallow feature extraction parameters for different types of veterinary herbal components (such as root, leaf, and flower herbs) to ensure feature targeting. Finally, it outputs a set of shallow microscopic feature vectors (one shallow feature vector corresponding to each effective area image).

[0060] The deep feature extraction unit, based on the set of shallow microscopic feature vectors output by the shallow feature extraction unit and combined with the set of effective region microscopic images output by the microscopic image acquisition and preprocessing module, employs an improved convolutional neural network (CNN) model (combining the advantages of ResNet and MobileNet, optimizing the network structure, and reducing computational load) to perform deep semantic feature extraction. During model training, the set of shallow feature vectors output by the shallow feature extraction unit is introduced as auxiliary input to guide the model to focus on key microscopic features, improving the accuracy of deep feature extraction. Simultaneously, the deep semantic features include the fine structural features of the Chinese herbal medicine components (such as cell wall thickness, starch grain morphology, and calcium oxalate crystal type) and the inter-component correlation features, and output a set of deep semantic feature vectors (one deep feature vector corresponding to each effective region image).

[0061] The feature fusion unit, based on the set of shallow microscopic feature vectors output by the shallow feature extraction unit and the set of deep semantic feature vectors output by the deep feature extraction unit, adopts an attention mechanism fusion algorithm to adaptively fuse the shallow microscopic features and deep semantic features. During the fusion process, attention weight allocation is used to highlight features that play an important role in identification (such as the calcium oxalate crystal feature of licorice and the stone cell feature of Coptis chinensis) and suppress the interference of invalid features. At the same time, the dimensionality of the fused feature vectors is reduced to remove redundant information, reduce the computational load of the subsequent identification module, and outputs a set of fused feature vectors (one fused feature vector corresponds to each effective region image).

[0062] Please see Figure 4 The veterinary Chinese medicine component identification module includes:

[0063] The component recognition model training unit constructs a sample library of veterinary Chinese medicine components containing standard microscopic images and corresponding feature vectors of veterinary Chinese medicines (such as Coptis chinensis, Scutellaria baicalensis, Glycyrrhiza uralensis, Pulsatilla chinensis, Lonicera japonica, etc.). Based on the fused feature vector set output by the microscopic feature extraction module, and combined with the standard data in the sample library, the component recognition model (a fusion model of improved support vector machine SVM and deep learning model) is trained. During the training process, a cross-validation mechanism is introduced to optimize the model parameters, improve the recognition accuracy and generalization ability of the model, avoid overfitting, and output the trained component recognition model and the veterinary Chinese medicine component sample library.

[0064] The component preliminary identification unit, based on the component identification model trained by the component identification model training unit and the fused feature vector set output by the microscopic feature extraction module, inputs the fused feature vector into the component identification model for preliminary identification. It determines the type of Chinese medicine component corresponding to each effective region image (e.g., identified as Coptis chinensis, Scutellaria baicalensis, etc.) and calculates the identification confidence score (range 0-1, confidence score ≥ 0.8 is considered valid identification). Identification results with confidence scores < 0.8 are marked for subsequent secondary identification and output a preliminary component identification result table (including image number, identified component type, and identification confidence score) and a set of low-confidence identified images (images with confidence scores < 0.8 and their corresponding feature vectors).

[0065] The secondary component identification and verification unit, based on the preliminary component identification result table and low-confidence identification image set output by the preliminary component identification unit, and combining the fusion feature vector set output by the microscopic feature extraction module and the veterinary Chinese medicine component sample library output by the component identification model training unit, performs secondary identification using feature comparison and manual verification. For low-confidence images with high similarity between features and standard component features, the identification model parameters are automatically adjusted, and the identification and confidence are recalculated. For images with large feature differences, manual verification prompts are output, and confirmation is made based on human experience. Simultaneously, the preliminary identification results and secondary identification results are integrated, erroneous identification results are eliminated, the final component identification results are determined, and a component identification result table (including formula sample number, type of each component, identification confidence, and verification method) is output.

[0066] Please see Figure 5 The prescription identification and quality assessment module includes:

[0067] The formula compatibility standard library construction unit collects standard compatibility information for common veterinary Chinese medicine formulas (such as Huanglian Jiedu San, Baitouweng Tang, and Gancao Xiexin Tang), including: formula name, standard component composition (including the types of Chinese medicines included), standard content range of each component, component ratio requirements, and quality qualification standards (such as component purity ≥95%, absence of toxic impurities, etc.). The compatibility information is standardized and organized to construct a veterinary Chinese medicine formula compatibility standard library, supporting subsequent identification and evaluation queries. At the same time, a standard library update mechanism is established to update the information in the library in real time based on new veterinary Chinese medicine formula standards and newly discovered characteristics of Chinese medicine components.

[0068] The formula authenticity identification unit, based on the veterinary Chinese medicine formula compatibility standard library constructed by the formula compatibility standard library construction unit and the component identification result table output by the veterinary Chinese medicine component identification module, performs formula authenticity identification. First, it compares the identified component types with the standard component types of the corresponding formula in the compatibility standard library to determine whether there are missing or redundant components (such as counterfeit products or impurities). Second, it compares the identification confidence of each component to determine whether there are cases where the content of effective components is too low or the content of counterfeit components is too high. If the identified component types and contents meet the standards, it is determined to be a qualified formula; otherwise, it is determined to be an unqualified formula, and the reason for unqualification is marked (such as missing Coptis chinensis component, adulterated Astragalus membranaceus, component content not meeting the standard), and the formula authenticity identification result table is output (including sample number, formula name, identification result, and reason for unqualification).

[0069] The formula quality grade assessment unit, based on the formula authenticity identification result table output by the formula authenticity identification unit and the component identification result table output by the veterinary Chinese medicine component identification module, and combined with the veterinary Chinese medicine formula compatibility standard library constructed by the formula compatibility standard library construction unit, assesses the quality grade of qualified formulas. First, quality assessment indicators are set, including: component purity (average confidence level of each component identification), component ratio deviation (difference between actual ratio and standard ratio), and effective component content (sum of confidence levels of core effective components). Then, based on the assessment indicator scores, qualified formulas are divided into two grades: Grade 1 (high-quality) and Grade 2 (qualified). Unqualified formulas are not included in the grade assessment. Simultaneously, quality improvement suggestions (such as component ratio adjustment and purity optimization) are output, along with a formula quality grade assessment result table (including sample number, formula name, quality grade, and scores for each assessment indicator).

[0070] Please see Figure 6 The anomaly warning and tracing module includes:

[0071] The anomaly warning unit, based on the formula authenticity identification result table and formula quality grade assessment result table output by the formula identification and quality assessment module, sets anomaly warning threshold (e.g., the proportion of unqualified formulas ≥5%, the proportion of secondary formulas ≥80%, the presence of toxic impurities), and monitors the identification and assessment results in real time. When anomalies exceeding the warning threshold are detected (e.g., the unqualified rate of a batch of formulas reaches 10%, or Xanthium sibiricum is detected as a counterfeit in a formula), it automatically issues the corresponding type of warning signal (audio warning, text warning, pop-up warning) and marks the warning level (general, severe, urgent). At the same time, it analyzes the cause of the anomaly (combining component identification results and sample basic information) and outputs an anomaly warning information table (including warning time, warning type, warning level, cause of anomaly, and sample number involved).

[0072] The full-process traceability unit, based on the abnormal warning information table output by the abnormal warning unit, the effective area microscopic image set output by the microscopic image acquisition and preprocessing module, and the formula authenticity identification result table output by the formula identification and quality assessment module, combined with the formula compatibility standard, performs full-process traceability of the formula. Through sample number association, it can query the formula's collection information, preprocessing parameters, feature extraction results, component identification process, identification and assessment results, clarify the source of abnormal formula problems (such as contamination in the sample collection process, adulteration in the production process, and deterioration in the storage process), and output a full-process traceability report of the formula (including key information and results of each step) and a traceability query record table (recording the query personnel, query time, and query content).

[0073] Please see Figure 7 The model adaptive optimization and intelligent application module for identification results includes:

[0074] The real-time performance monitoring unit for the identification model monitors the running performance of the model (improved convolutional neural network (CNN) model and component identification model) in real time, based on the component identification result table output by the veterinary Chinese medicine component identification module and the formula authenticity identification result table and formula quality grade evaluation result table output by the formula identification and quality assessment module. Simultaneously, it sets performance evaluation indicators, including: component identification accuracy, formula identification accuracy, model running efficiency (time spent processing a single sample), and anomaly identification sensitivity. It compares the current performance indicators with preset standard thresholds in real time, records performance fluctuation data, identifies scenarios of model performance decline (such as accuracy below 90% or running time exceeding 10 minutes), and outputs a real-time model performance monitoring report (including various performance indicators and fluctuation trends), a performance anomaly warning signal (output only when the indicator exceeds the threshold), and a preliminary analysis table of the causes of model performance fluctuations (combined with identification results and feature extraction data).

[0075] The model adaptive optimization unit, based on the real-time performance monitoring report, performance anomaly warning signal, and preliminary analysis table of model performance fluctuations output by the real-time performance monitoring unit of the identification model, and combined with the fusion feature vector set and prescription compatibility standards output by the microscopic feature extraction module, performs adaptive optimization of the model. For performance degradation, it automatically adjusts model parameters: if incomplete feature extraction leads to a decrease in accuracy, it optimizes the convolution kernel parameters and attention weights of the deep feature extraction model; if changes in component characteristics (such as changes in microscopic features due to differences in the origin of Chinese medicinal herbs) cause recognition deviations, it automatically updates the training samples of the component recognition model (supplementing new feature samples) and readjusts the model. Furthermore, during the optimization process, it retains historically optimal model parameters, supports rollback operations, and outputs the optimized model and an optimized model performance prediction report.

[0076] The intelligent transformation unit for identification results, based on the optimized model performance prediction report output by the model adaptive optimization unit, combined with the formula authenticity identification result table and formula quality grade evaluation result table output by the formula identification and quality assessment module, and the abnormal warning information table output by the abnormal warning and traceability module, transforms the identification results into a feasible intelligent application solution, adaptable to different application scenarios. For veterinary Chinese medicine manufacturers, it outputs production optimization suggestions (such as adjusting the formula component ratio and optimizing raw material screening standards); for livestock breeding bases, it outputs medication guidance solutions (such as adjusting the dosage according to the formula quality grade and recommending suitable diseases); for regulatory agencies, it outputs key quality supervision reminders (such as focusing on investigating the production source of a batch of unqualified formulas); at the same time, it standardizes and organizes the identification results and application solutions to form an exportable report template and outputs a feasibility analysis table for the implementation of the application solution.

[0077] The optimization model verification and iteration unit, based on the optimized model output by the model adaptive optimization unit and the application scheme feasibility analysis table output by the intelligent conversion unit of the identification results, and combined with the set of effective area microscopic images of the new batch (randomly selected samples) output by the microscopic image acquisition and preprocessing module, inputs the new batch of sample data into the optimized model, compares the verification results with the manual identification results, calculates the actual performance index of the optimized model, and determines whether it meets the preset standard; at the same time, it collects feedback data after the application scheme is implemented (such as the effect of ratio adjustment in production enterprises and the effect of drug use in breeding bases), and transforms the feedback data into supplementary basis for model optimization; if the verification is successful, the optimized model is set as the current model used by the system, and the component standard feature library is updated; if the verification fails, it returns to unit 6.2 to readjust the optimization parameters.

[0078] In practical use, the specific steps are as follows:

[0079] S1, Microscopic Image Acquisition and Preprocessing: Acquire microscopic images of veterinary Chinese herbal medicine prescriptions, and output a set of effective region microscopic images through noise removal, standardization and segmentation operations;

[0080] S2, Microscopic feature extraction: Based on the set of effective area microscopic images output by the microscopic image acquisition and preprocessing steps, extract the shallow microscopic features and deep semantic features of each Chinese herbal medicine component in the prescription, and output a set of fused feature vectors through feature fusion.

[0081] S3, Identification of components of veterinary Chinese medicine: Based on the fusion feature vector set output by the microscopic feature extraction step, identify each Chinese medicine component in the veterinary Chinese medicine prescription, determine the type and purity of the components, and output a component identification result table;

[0082] S4, Formula Identification and Quality Assessment: Based on the component identification result table output by the veterinary Chinese medicine component identification step, combined with the compatibility standard of veterinary Chinese medicine formula, the authenticity of the formula and the quality grade assessment are realized, and the authenticity identification result table and the quality grade assessment result table of the formula are output.

[0083] S5, Anomaly Warning and Traceability: Based on the formula authenticity identification result table and formula quality grade assessment result table output by the formula identification and quality assessment steps, identify abnormal situations of the formula and issue warnings. At the same time, realize the traceability of the formula throughout the whole process and output the anomaly warning information table and the formula traceability report throughout the whole process.

[0084] S6, Model Adaptive Optimization and Intelligent Application of Identification Results: Based on the component identification result table output by the veterinary Chinese medicine component identification step and the formula authenticity identification result table and formula quality grade evaluation result table output by the formula identification and quality evaluation step, the deep learning identification model is dynamically and adaptively optimized, and the identification and quality evaluation results are transformed into a feasible intelligent application solution.

[0085] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A deep learning-based microscopic image identification system for veterinary traditional Chinese medicine prescriptions, characterized in that, include: The microscopic image acquisition and preprocessing module acquires microscopic images of veterinary Chinese medicine prescriptions and outputs a set of effective region microscopic images through noise removal, standardization, and segmentation operations. The microscopic feature extraction module includes a shallow feature extraction unit, a deep feature extraction unit, and a feature fusion unit. First, the shallow feature extraction unit extracts shallow microscopic features from the effective region microscopic image set output by the microscopic image acquisition and preprocessing module, and outputs a shallow microscopic feature vector set. Next, the deep feature extraction unit, based on the shallow microscopic feature vector set output by the shallow feature extraction unit and combined with the effective region microscopic image set output by the microscopic image acquisition and preprocessing module, uses an improved convolutional neural network model to extract deep semantic features, and outputs a deep semantic feature vector set. Finally, the feature fusion unit, based on the shallow microscopic feature vector set output by the shallow feature extraction unit and the deep semantic feature vector set output by the deep feature extraction unit, uses an attention mechanism fusion algorithm to adaptively fuse the shallow microscopic features and deep semantic features; simultaneously, it performs dimensionality reduction processing on the fused feature vectors and outputs a fused feature vector set. The veterinary Chinese medicine component identification module identifies each Chinese medicine component in the veterinary Chinese medicine formula based on the fusion feature vector set output by the microscopic feature extraction module, determines the type and purity of the components, and outputs a component identification result table. The formula identification and quality assessment module, based on the component identification result table output by the veterinary Chinese medicine component identification module, combined with the compatibility standards of veterinary Chinese medicine formulas, realizes the identification of the authenticity of the formula and the assessment of its quality grade, and outputs the formula authenticity identification result table and the formula quality grade assessment result table. The anomaly warning and traceability module, based on the formula authenticity identification result table and formula quality grade assessment result table output by the formula identification and quality assessment module, identifies abnormal situations of the formula and issues warnings, while realizing full-process traceability of the formula and outputting an anomaly warning information table and a full-process traceability report of the formula. The model adaptive optimization and intelligent application module for identification results, based on the component identification result table output by the veterinary Chinese medicine component identification module and the formula authenticity identification result table and formula quality grade evaluation result table output by the formula identification and quality evaluation module, performs dynamic adaptive optimization of the deep learning identification model, and transforms the identification and quality evaluation results into a feasible intelligent application solution.

2. The deep learning-based microscopic image identification system for veterinary traditional Chinese medicine prescriptions according to claim 1, characterized in that, The microscopic image acquisition and preprocessing module includes: The microscopic image acquisition unit, based on the sample characteristics of veterinary Chinese medicine prescriptions, sets standardized acquisition parameters to acquire images of prescription samples from different batches and different parts in all directions, and outputs a set of original microscopic images. The image noise removal unit identifies various types of noise in the image based on the original set of microscopic images output by the microscopic image acquisition unit. At the same time, it adopts an adaptive noise removal algorithm to set differentiated filtering parameters for different types of noise. While removing noise, it retains the subtle features in the microscopic image and outputs a set of denoising microscopic images. The image normalization unit performs normalization processing on the set of denoised microscopic images output by the image noise removal unit; at the same time, it performs geometric correction on the images and outputs a set of normalized microscopic images. The image segmentation unit, based on the standardized set of microscopic images output by the image standardization unit, uses a deep learning semantic segmentation algorithm to automatically segment the microscopic images, separating the effective discriminative regions from the background regions in the images, and outputting a set of effective region microscopic images.

3. The deep learning-based microscopic image identification system for veterinary traditional Chinese medicine prescriptions according to claim 1, characterized in that, The veterinary traditional Chinese medicine component identification module includes: The component identification model training unit constructs a sample library of veterinary Chinese medicine components containing standard microscopic images and corresponding feature vectors. Based on the fusion feature vector set output by the microscopic feature extraction module, combined with the standard data in the sample library, the component identification model is trained, and the trained component identification model and the veterinary Chinese medicine component sample library are output. The component preliminary identification unit, based on the component identification model trained by the component identification model training unit and the fused feature vector set output by the microscopic feature extraction module, inputs the fused feature vector into the component identification model to perform preliminary identification, determines the type of Chinese medicine component corresponding to each effective region image, calculates the identification confidence, marks the identification results of the confidence, and outputs a preliminary component identification result table and a set of low-confidence identification images. The secondary component identification and verification unit, based on the preliminary component identification result table and low-confidence identification image set output by the preliminary component identification unit, and combined with the fusion feature vector set output by the microscopic feature extraction module and the veterinary Chinese medicine component sample library output by the component identification model training unit, performs secondary identification by feature comparison and manual verification; at the same time, it integrates the preliminary identification results and the secondary identification results and outputs the component identification result table.

4. The deep learning-based microscopic image identification system for veterinary traditional Chinese medicine prescriptions according to claim 1, characterized in that, The prescription identification and quality assessment module includes: The formula compatibility standard library construction unit collects standard compatibility information of veterinary Chinese medicine formulas, and standardizes and organizes the compatibility information to construct a veterinary Chinese medicine formula compatibility standard library; The formula authenticity identification unit, based on the veterinary Chinese medicine formula compatibility standard library constructed by the formula compatibility standard library construction unit and the component identification result table output by the veterinary Chinese medicine component identification module, performs formula authenticity identification and outputs formula authenticity identification result table. The formula quality grade assessment unit, based on the formula authenticity identification result table output by the formula authenticity identification unit and the component identification result table output by the veterinary Chinese medicine component identification module, and combined with the veterinary Chinese medicine formula compatibility standard library constructed by the formula compatibility standard library construction unit, assesses the quality grade of qualified formulas and outputs a formula quality grade assessment result table.

5. The deep learning-based microscopic image identification system for veterinary traditional Chinese medicine prescriptions according to claim 1, characterized in that, The anomaly warning and source tracing module includes: The anomaly warning unit, based on the formula authenticity identification result table and formula quality grade evaluation result table output by the formula identification and quality evaluation module, sets an anomaly warning threshold and monitors the identification and evaluation results in real time. When an anomaly exceeding the warning threshold is detected, it automatically issues a warning signal of the corresponding type and marks the warning level. At the same time, it analyzes the cause of the anomaly and outputs an anomaly warning information table. The full-process traceability unit, based on the abnormal warning information table output by the abnormal warning unit, the set of effective area microscopic images output by the microscopic image acquisition and preprocessing module, and the formula authenticity identification result table output by the formula identification and quality assessment module, combined with the formula compatibility standard, performs full-process traceability of the formula and outputs a full-process traceability report and a traceability query record table.

6. The deep learning-based microscopic image identification system for veterinary traditional Chinese medicine prescriptions according to claim 1, characterized in that, The intelligent application module for model adaptive optimization and identification results includes: The real-time performance monitoring unit for the identification model monitors the model's performance in real time based on the component identification result table output by the veterinary Chinese medicine component identification module and the formula authenticity identification result table and formula quality grade evaluation result table output by the formula identification and quality assessment module. Simultaneously, it sets performance evaluation indicators, compares the current performance indicators with preset standard thresholds in real time, records performance fluctuation data, identifies scenarios of model performance decline, and outputs a real-time model performance monitoring report, performance anomaly warning signal, and preliminary analysis table of the causes of model performance fluctuations. The model adaptive optimization unit, based on the real-time performance monitoring report, performance anomaly warning signal and preliminary analysis table of model performance fluctuation reasons output by the real-time performance monitoring unit of the identification model, combined with the fusion feature vector set and prescription compatibility standard output by the microscopic feature extraction module, performs adaptive optimization of the model and outputs the optimized model and the optimized model performance prediction report. The intelligent transformation unit for identification results, based on the optimized model performance prediction report output by the model adaptive optimization unit, combined with the formula authenticity identification result table and formula quality grade evaluation result table output by the formula identification and quality assessment module, and the abnormal warning information table output by the abnormal warning and traceability module, transforms the identification results into a feasible intelligent application solution; at the same time, it standardizes and organizes the identification results and application solutions to form an exportable report template, and outputs a feasibility analysis table for the implementation of the application solution; The optimization model verification and iteration unit, based on the optimized model output by the model adaptive optimization unit and the application scheme feasibility analysis table output by the intelligent conversion unit of the identification results, combined with the new batch of effective area microscopic images output by the microscopic image acquisition and preprocessing module, compares the verification results with the manual identification results, calculates the actual performance index of the optimized model, and determines whether it meets the preset standard; at the same time, it collects feedback data after the application scheme is implemented.

Images