Training method of southern semen cuscutae grating cell intelligent detection model

Abstract

The invention belongs to the technical field of traditional Chinese medicinal material microscopic imaging intelligent processing and analysis, and discloses a southern semen cuscutae grating cell intelligent detection model training method, which comprises the following steps: collecting digital microscopic images of southern semen cuscutae grating cells, each digital microscopic image having a feature label; irregular aggregation and splitting are carried out on digital microscopic imaging based on feature labeling, and a training cluster and an optimization cluster are obtained; constructing an initial southern semen cuscutae grating cell intelligent detection model based on a deep learning technology; performing cross training on the initial southern semen cuscutae grating cell intelligent detection model by utilizing the training cluster to obtain an optimal southern semen cuscutae grating cell intelligent detection model; performing iterative optimization on the optimal southern semen cuscutae grating cell intelligent detection model by utilizing the optimization cluster to obtain a target southern semen cuscutae grating cell intelligent detection model; the precision and the efficiency of Chinese herbal medicine quality inspection personnel in the southern semen cuscutae grid-like cell detection and screening work are greatly improved.

Description

Technical Field

[0001] The present application relates to the technical field of intelligent processing and analysis of microscopic imaging of traditional Chinese medicine, and in particular to a method for training an intelligent detection model of palisade cells of Cuscuta australis. Background Art

[0002] With the sweeping global information wave, intelligence has become a major trend in the development of various industries. At present, the supervision of traditional Chinese medicine and ethnic medicine is constantly increasing, the inspection tasks are increasing year by year, and there is an extreme shortage of traditional identification talents. It is urgent to carry out research on intelligent identification methods of traditional Chinese medicine and ethnic medicine based on digital images of traditional Chinese medicine specimens and apply cutting-edge technologies such as computer image recognition to fill the gaps in the industry and lay the foundation for the development of new scientific supervision tools.

[0003] Authenticity identification of Chinese medicinal materials is an important guarantee for the safety and effectiveness of Chinese medicine; however, with the continuous increase in the supervision of Chinese medicine, the workload of Chinese medicine inspection is also gradually increasing. Therefore, a fast and efficient work pace has become the daily routine of Chinese medicine inspectors; currently, there is an extreme shortage of talents in authenticity identification of Chinese medicinal materials. The mastery of authenticity identification of Chinese medicinal materials needs to be based on long-term practical accumulation and accurate observation of a large number of accurate samples. It is urgent to carry out key technology research on the intelligent application of Chinese medicine ethnic specimens to serve daily inspection and grassroots supervision.

[0004] Compared with traditional manual identification, the use of artificial intelligence detection models can greatly reduce the manpower required for identification work, improve the work efficiency of identification personnel, and solve the problems of shortage of identification talents and high work pressure of traditional Chinese medicine inspectors; this method utilizes digital microscopic imaging of the palisade cells of southern dodder, applies image recognition technologies such as convolutional networks, and obtains a deep convolutional neural network model with high reliability and good generalization performance by training feature extraction and feature recognition modules, and identifies the palisade cells of southern dodder in the application.

[0005] In view of this, the present application is proposed to provide a method for training an intelligent detection model of palisade cells of Cuscuta australis that overcomes the above problems or at least partially solves the above problems. Summary of the invention

[0006] In order to solve the above problems, the present invention discloses a method for training a palisade cell intelligent detection model of Cuscuta australis, comprising:

[0007] S1. Collect digital microscopic images of palisade cells of Cuscuta australis, each of which has feature annotations; perform random aggregation on the digital microscopic images based on the feature annotations to obtain A microscopic imaging clusters, split the microscopic imaging clusters by clusters, and obtain training clusters and optimization clusters;

[0008] S2. Constructing the initial intelligent detection model of palisade cells of Cuscuta australis based on deep learning technology;

[0009] S3, using the training cluster to cross-train the initial southern dodder palisade cell intelligent detection model to obtain the most suitable southern dodder palisade cell intelligent detection model;

[0010] S4. Use the optimization cluster to iteratively optimize the most suitable southern dodder palisade cell intelligent detection model to obtain the target southern dodder palisade cell intelligent detection model.

[0011] Furthermore, the feature annotation includes: feature classification annotation and target positioning annotation; target positioning annotation includes: marking position coordinates on digital microscopic imaging; feature classification annotation includes: marking image features of palisade cells of southern dodder on digital microscopic imaging.

[0012] Furthermore, the method of randomly aggregating digital microscopic imaging includes:

[0013] Based on feature annotation, a microscopic feature vector is constructed. The similarity between different microscopic feature vectors is calculated using a similarity calculation formula. Based on the similarity, a similarity matrix M is constructed. ij represents the element in the i-th row and j-th column of the similarity matrix, then M ij represents the similarity between the i-th microscopic feature vector and the j-th microscopic feature vector; for the i-th row of the similarity matrix, the indexes of the columns whose element values ​​are greater than the preset similarity threshold constitute the neighboring set of the i-th microscopic feature vector; the fitness of each microscopic feature vector is calculated based on the similarity matrix and the neighboring set, and the formula for fitness calculation is: Where D(i) represents the local fitness of the i-th microscopic feature vector, N represents the neighboring set, Sim represents the size of the neighboring set, and σ represents the adjustment parameter. Cluster analysis is performed through local fitness to classify similar digital microscopic images into one category, and A microscopic imaging clusters are obtained, where A is an integer greater than zero.

[0014] Furthermore, the construction method of the initial Cuscuta australis palisade cell intelligent detection model includes:

[0015] The initialization feature extraction model is a neural network model. The network layer structure includes: input layer, convolution layer, pooling layer, fully connected layer and output layer. The Sigmoid function is used as the activation function. The input layer uses the training set as input, and the output layer uses the feature cell map as output.

[0016] The feature recognition model is initialized as the Faster R-CNN model, and the region proposal network is defined to use the Anchor mechanism as the candidate box generation mechanism. The classifier is the SoftMax classifier, the regressor is the Lasso regressor, the feature cell map is used as the input of the feature recognition model, and the target recognition result is used as the output of the feature recognition model.

[0017] Furthermore, the method of cross-training the initial Cuscuta australis palisade cell intelligent detection model includes:

[0018] Based on the total number of clusters K in the training clusters, K model trainings are performed. Each time the model training takes one of the training clusters as a validation set, and the other training clusters are trained as training sets to obtain K initial southern dodder palisade cell intelligent detection models; the initial southern dodder palisade cell intelligent detection model includes a feature extraction model and a feature recognition model. For each model training, the digital microscopic imaging in the training set is used as the input of the feature extraction model, and the digital microscopic imaging in the training set is convolved through a convolution layer to obtain a feature map; the feature map is pooled through a pooling layer to obtain a low-dimensional feature map; the low-dimensional feature map is feature mapped through a fully connected layer to obtain a feature cell map; the feature recognition model is used to perform feature recognition on the feature cell map to obtain a target recognition result, and the target recognition result is confidence calculated to obtain a confidence parameter; the average value of the confidence parameters obtained in each training is calculated, and the initial southern dodder palisade cell intelligent detection model with the largest average value is used as the most suitable southern dodder palisade cell intelligent detection model.

[0019] Furthermore, the method of performing the convolution operation on the digital microscopic imaging in the training set includes: presetting the moving stride of the convolution kernel and the size of the convolution matrix, the convolution kernel of the convolution layer slides in the digital microscopic imaging with the moving stride, and after each sliding, multiplying and summing the pixel elements of the selected digital microscopic imaging area that is consistent with the size of the convolution matrix with the convolution kernel to extract the local features of the digital microscopic imaging;

[0020] The method of performing a pooling operation on the feature map includes: dynamically adjusting the pooling window based on the size of the acquired feature map and the size of the digital microscopic imaging, and the length calculation formula of the pooling window is: Among them, HEI represents the length of the pooling window, H in Represents the length of digital microscopy imaging, H out Represents the length of the feature map, Floor() represents the floor function; the width of the pooling window is calculated as: Among them, WID represents the width of the pooling window, W in Represents the width of the digital microscope image, W outRepresents the width of the feature map; select elements in the feature map according to the length and width of the pooling window. When the length of the pooling window is equal to the width, take the average value of the elements in the pooling window. When the length of the pooling window is not equal to the width, take the maximum value of the elements in the pooling window.

[0021] Furthermore, the method of performing feature recognition on the characteristic cell map includes:

[0022] The region proposal network scans the feature cell map and generates candidate boxes of different scales and aspect ratios using the Anchor mechanism. The position and size of the bounding box are adjusted through the regressor. The object of each candidate box and the category to which the object belongs are identified through the classifier. The candidate boxes identified with the same category are spliced ​​to obtain a new candidate box. According to the size of the feature cell map and the size of the new candidate box, the area selected by the candidate box is converted into regional data, and the converted regional data is the target recognition result.

[0023] Furthermore, the formula for calculating the confidence of the target recognition result is:

[0024] Where C(x) represents the confidence parameter of the x-th target recognition result, All represents the total number of categories of the recognized content in the target recognition result, and S represents the number of categories in the target recognition result that are consistent with the feature annotation content. represents the correlation between the x-th target recognition result and the e-th category in the x-th feature annotation, β e Represents the weight value of the e-th category in the target recognition result, x e represents the e-th category of the x-th target recognition result, Represents the e-th category of the x-th feature annotation.

[0025] Furthermore, the method of iteratively optimizing the optimal palisade cell intelligent detection model of Cuscuta australis includes:

[0026] A genetic algorithm is used to perform convolution optimization on the convolution layer in the feature extraction model to obtain the optimal convolution group. The optimal convolution group is used as the convolution kernel of the convolution layer in the feature extraction model to obtain the target feature extraction model. The digital microscopic imaging in the optimized cluster is used as the input of the target feature extraction model for feature extraction to obtain feature cell maps corresponding to different optimized clusters. The cell map features corresponding to the same optimized cluster constitute a feature map cluster. The feature map cluster is used as the input of the feature recognition model for feature recognition to obtain the target recognition result. The target recognition result is trained by a comprehensive loss function to supervise the entire learning process. The non-maximum suppression algorithm is used to suppress the repeated results of different candidate boxes to obtain the target feature recognition model.

[0027] Furthermore, the method of performing convolution optimization on the convolution layer in the feature extraction model includes:

[0028] For the convolution layer in the feature extraction model, P initial convolution kernel groups are preset, each convolution kernel group has E convolution kernels, the optimal convolution group is initialized to be empty, the weight parameters in each convolution kernel are initialized, and the weight parameters are randomly taken based on the normal distribution L(a, b); based on the convolution kernel in the convolution layer, the digital microscopic imaging in the optimized cluster is used as the input of the feature extraction model for forward propagation, the input of the previous convolution layer is used as the output of the next convolution layer, the cross entropy function is used as the loss function of each convolution layer, the convolution kernel group is traversed, each convolution kernel group is used as the convolution kernel in the convolution layer, and the inverse of the loss function value is used as the convolution fitness of each convolution kernel group; the selection probability of each convolution kernel group is calculated based on the convolution fitness, and the calculation formula of the selection probability is: Among them, Value(c) represents the selection probability of the cth convolution kernel group, d represents the index of the convolution kernel group, Fit(d) represents the convolution fitness of the dth convolution kernel group, and Fit(c) represents the convolution fitness of the cth convolution kernel group; iterative selection is performed based on the selection probability, and F convolution kernel groups are selected as the parent convolution kernel groups each time. The parent convolution kernel groups are combined in pairs using the permutation and combination method to obtain the parent convolution pairs. For each pair of parent convolution pairs, the convolution kernel with the largest modulus value in the parent convolution kernel group is selected as the crossover node for single-point crossover to obtain the child convolution kernel group; the increment value is preset, the convolution kernels in the child convolution kernel group are traversed, and a random selection algorithm is used. The weight parameters in the convolution kernel are selectively mutated, and each weight parameter that is selectively mutated is added to the incremental value to obtain a mutated convolution kernel group; the mutated convolution kernel group is used as the convolution kernel in the convolution layer, and the convolution fitness of the mutated convolution kernel group is calculated. When the optimal convolution group is empty, the mutated convolution kernel group with the largest convolution fitness value is selected as the new optimal convolution group. When the optimal convolution group is not empty, the mutated convolution kernel group with the largest convolution fitness value that is greater than the convolution fitness value of the optimal convolution group is selected as the new optimal convolution group. Repeat until the convolution fitness of the optimal convolution group no longer changes, and output the optimal convolution group at this time as the convolution kernel of the convolution layer in the feature extraction model.

[0029] The technical effects and advantages of the intelligent detection model training method of palisade cells of southern dodder of the present invention are as follows:

[0030] The present invention utilizes artificial intelligence technologies such as deep learning to assist Chinese herbal medicine quality inspectors in conducting intelligent detection of palisade cells of southern dodder and improve the accuracy and efficiency of screening work, and has significant application prospects; by combining deep learning technology with knowledge of Chinese medicinal material identification and detection, a digital microscopic imaging-based intelligent detection technology for palisade cells of southern dodder is realized, which further improves the accuracy and efficiency of Chinese medicinal material identification and detection, effectively solves the defects in the detection process of palisade cells of southern dodder, and promotes the application and development of a new generation of artificial intelligence technology in intelligent detection and auxiliary identification of Chinese medicinal materials; through an efficient automated training process, the need for manual intervention is reduced, the level of intelligent detection is improved, and the needs of modern biological research and clinical applications are met. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the technical solution of the present application, the drawings required for use in the description of the present application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0032] Figure 1 This is a flowchart of the steps of a method for training a palisade cell intelligent detection model for Cuscuta australis provided in one embodiment of the present application;

[0033] Figure 2 This is a schematic diagram of a network structure of a feature extraction model of a palisade cell intelligent detection model of Cuscuta australis provided in one embodiment of the present application;

[0034] Figure 3 This is a schematic diagram of the network structure of a feature recognition model of an intelligent detection model for palisade cells of Cuscuta australis provided in one embodiment of the present application. DETAILED DESCRIPTION

[0035] In order to make the objects, features and advantages of the present application more obvious and understandable, the present application is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present application.

[0036] It should be noted that, in any embodiment of the present invention, a method for training a model for intelligent detection of palisade cells of southern dodder is applied to intelligent detection of palisade cells of southern dodder through digital microscopic imaging of palisade cells of southern dodder. Confidence parameters, including the confidence interval of a probability sample, are interval estimates of a certain population parameter of this sample. The confidence interval shows the degree to which the true value of this parameter has a certain probability of falling around the measurement result. The confidence interval gives the range of credibility of the measured value of the measured parameter, that is, the "certain probability". This probability is called a confidence parameter.

[0037] Reference Figure 1 , shows a method for training a palisade cell intelligent detection model for Cuscuta australis provided in an embodiment of the present application, comprising:

[0038] S1. Collect digital microscopic images of palisade cells of Cuscuta australis, each of which has feature annotations; perform random aggregation on the digital microscopic images based on the feature annotations to obtain A microscopic imaging clusters, split the microscopic imaging clusters by clusters, and obtain training clusters and optimization clusters;

[0039] S2. Constructing the initial intelligent detection model of palisade cells of Cuscuta australis based on deep learning technology;

[0040] S3, using the training cluster to cross-train the initial southern dodder palisade cell intelligent detection model to obtain the most suitable southern dodder palisade cell intelligent detection model;

[0041] S4. Use the optimization cluster to iteratively optimize the most suitable southern dodder palisade cell intelligent detection model to obtain the target southern dodder palisade cell intelligent detection model.

[0042] The feature annotation includes feature classification annotation and target positioning annotation of the microscopic observation morphologies of three types of palisade cells of southern dodder; wherein, the feature classification annotation includes the image features of the microscopic observation morphologies of the three types of palisade cells of southern dodder on the digital microscopic imaging; the target positioning annotation includes the target positioning annotation of the image features on the digital microscopic imaging; wherein, the microscopic observation morphologies of the three types of palisade cells of southern dodder include: surface view of palisade cells of southern dodder, complete side view of palisade cells of southern dodder, and incomplete side view of palisade cells of southern dodder.

[0043] Methods for random aggregation of digital microscopy images include:

[0044] Based on feature annotation, a microscopic feature vector is constructed. The similarity between different microscopic feature vectors is calculated using a similarity calculation formula. Based on the similarity, a similarity matrix M is constructed. ij represents the element in the i-th row and j-th column of the similarity matrix, then Mij represents the similarity between the i-th microscopic feature vector and the j-th microscopic feature vector; for the i-th row of the similarity matrix, the indexes of the columns whose element values ​​are greater than the preset similarity threshold constitute the neighboring set of the i-th microscopic feature vector; the fitness of each microscopic feature vector is calculated based on the similarity matrix and the neighboring set, and the formula for fitness calculation is: Where D(i) represents the local fitness of the i-th microscopic feature vector, N represents the neighboring set, Sim represents the size of the neighboring set, and σ represents the adjustment parameter, which is used to control the sensitivity of the local fitness. Cluster analysis is performed through local fitness to classify similar digital microscopic images into one category, and A microscopic imaging clusters are obtained, where A is an integer greater than zero. Commonly used clustering analysis methods include K-means clustering, hierarchical clustering, and density clustering. By performing feature annotation and cluster analysis on the digital microscopic images of palisade cells of Cuscuta australis, different types of palisade cells can be more accurately identified and detected. Feature annotation includes classification of microscopic observation morphology and target positioning, which enables the model to effectively distinguish cell types during the recognition process and improve the accuracy of detection.

[0045] The initialization feature extraction model is a neural network model. The network layer structure includes: input layer, convolution layer, pooling layer, fully connected layer and output layer. The Sigmoid function is used as the activation function. The input layer uses the training set as input, and the output layer uses the feature cell map as output.

[0046] The feature recognition model is initialized as the Faster R-CNN model, and the region proposal network is defined with the Anchor mechanism as the candidate box generation mechanism. The classifier is the SoftMax classifier, and the regressor is the Lasso regressor. The feature cell map is used as the input of the feature recognition model, and the target recognition result is used as the output of the feature recognition model. The region proposal network and the Anchor mechanism are used to generate candidate boxes of different scales and aspect ratios, which enhances the model's ability to recognize targets in complex environments. Through the adjustment of the regressor and the discrimination of the classifier, the model can effectively splice candidate boxes of the same category, thereby improving the detection accuracy and recall rate.

[0047] The cross-training methods for the initial southern dodder palisade cell intelligent detection model include:

[0048] Based on the total number of clusters K in the training clusters, K model trainings are performed. Each time the model training is performed, one of the training clusters is used as a validation set, and the other training clusters are used as training sets for training, so as to obtain K initial intelligent detection models for palisade cells of dodder. The initial intelligent detection model for palisade cells of dodder includes a feature extraction model and a feature recognition model. For each model training, the digital microscopic imaging in the training set is used as the input of the feature extraction model, and the digital microscopic imaging in the training set is convolved through a convolution layer to obtain a feature map; the feature map is pooled through a pooling layer to obtain a low-dimensional feature map; the low-dimensional feature map is feature mapped through a fully connected layer to obtain a feature cell map; the feature recognition model is used to perform feature recognition on the feature cell map to obtain a target recognition result, and a confidence calculation is performed on the target recognition result. The formula for the confidence calculation is:

[0049] Where C(x) represents the confidence parameter of the x-th target recognition result, All represents the total number of categories of the recognized content in the target recognition result, and S represents the number of categories in the target recognition result that are consistent with the feature annotation content. represents the correlation between the x-th target recognition result and the e-th category in the x-th feature annotation, β e Represents the weight value of the e-th category in the target recognition result, x e represents the e-th category of the x-th target recognition result, Represents the e-th category of the x-th feature annotation; calculates the average value of the confidence parameters obtained in each training, and takes the initial southern dodder palisade cell intelligent detection model with the largest average value as the most suitable southern dodder palisade cell intelligent detection model.

[0050] The method of performing convolution operation on the digital microscopic imaging in the training set includes: presetting the moving stride of the convolution kernel and the size of the convolution matrix, the convolution kernel of the convolution layer slides in the digital microscopic imaging with the moving stride, and after each slide, multiplying and summing the pixel elements of the selected digital microscopic imaging area with the same size of the convolution matrix with the convolution kernel to extract the local features of the digital microscopic imaging;

[0051] The method of performing pooling operation on the feature map includes: dynamically adjusting the pooling window based on the size of the acquired feature map and the size of the digital microscopic imaging, and the length calculation formula of the pooling window is: Among them, HEI represents the length of the pooling window, H in Represents the length of digital microscopy imaging, H out Represents the length of the feature map, Floor() represents the floor function; the width of the pooling window is calculated as: Among them, WID represents the width of the pooling window, W in Represents the width of the digital microscope image, W outRepresents the width of the feature map; select elements in the feature map according to the length and width of the pooling window. When the length of the pooling window is equal to the width, take the average value of the elements in the pooling window. When the length of the pooling window is not equal to the width, take the maximum value of the elements in the pooling window.

[0052] Methods for feature recognition of characteristic cell maps include:

[0053] The region proposal network scans the feature cell map and generates candidate boxes of different scales and aspect ratios using the Anchor mechanism. The position and size of the bounding box are adjusted through the regressor. The object of each candidate box and the category to which the object belongs are identified through the classifier. The candidate boxes identified with the same category are spliced ​​to obtain a new candidate box. According to the size of the feature cell map and the size of the new candidate box, the area selected by the candidate box is converted into regional data, and the converted regional data is the target recognition result.

[0054] The iterative optimization method of the optimal palisade cell intelligent detection model of southern dodder includes:

[0055] A genetic algorithm is used to perform convolution optimization on the convolution layer in the feature extraction model to obtain the optimal convolution group. The optimal convolution group is used as the convolution kernel of the convolution layer in the feature extraction model to obtain the target feature extraction model. The digital microscopic imaging in the optimized cluster is used as the input of the target feature extraction model for feature extraction to obtain feature cell maps corresponding to different optimized clusters. The cell map features corresponding to the same optimized cluster constitute a feature map cluster. The feature map cluster is used as the input of the feature recognition model for feature recognition to obtain the target recognition result. The target recognition result is trained by a comprehensive loss function to supervise the entire learning process. The non-maximum suppression algorithm is used to suppress the repeated results of different candidate boxes to obtain the target feature recognition model.

[0056] Ways to perform convolution optimization on the convolutional layers in the feature extraction model include:

[0057] For the convolution layer in the feature extraction model, P initial convolution kernel groups are preset, each convolution kernel group has E convolution kernels, the optimal convolution group is initialized to be empty, the weight parameters in each convolution kernel are initialized, and the weight parameters are randomly taken based on the normal distribution L(a, b); based on the convolution kernel in the convolution layer, the digital microscopic imaging in the optimized cluster is used as the input of the feature extraction model for forward propagation, the input of the previous convolution layer is used as the output of the next convolution layer, the cross entropy function is used as the loss function of each convolution layer, the convolution kernel group is traversed, each convolution kernel group is used as the convolution kernel in the convolution layer, and the inverse of the loss function value is used as the convolution fitness of each convolution kernel group; the selection probability of each convolution kernel group is calculated based on the convolution fitness, and the calculation formula of the selection probability is: Among them, Value(c) represents the selection probability of the c-th convolution kernel group, d represents the index of the convolution kernel group, Fit(d) represents the convolution fitness of the d-th convolution kernel group, and Fit(c) represents the convolution fitness of the c-th convolution kernel group; iterative selection is performed based on the selection probability, and F convolution kernel groups are selected as the parent convolution kernel groups each time. The parent convolution kernel groups are combined in pairs using the permutation and combination method to obtain parent convolution pairs. For each pair of parent convolution pairs, the convolution kernel with the largest modulus value in the parent convolution kernel group is selected as the crossover node for single-point crossover to obtain the child convolution kernel group; the increment value is preset, and the increment value is a random small value used to increase the diversity of the child convolution kernel group, and traverse the child convolution kernel group. The convolution kernel in the convolution kernel is selected and mutated by a random selection algorithm. Each weight parameter selected and mutated is added to the increment value to obtain a mutated convolution kernel group. The mutated convolution kernel group is used as the convolution kernel in the convolution layer, and the convolution fitness of the mutated convolution kernel group is calculated. When the optimal convolution group is empty, the mutated convolution kernel group with the largest convolution fitness value is selected as the new optimal convolution group. When the optimal convolution group is not empty, the mutated convolution kernel group with the largest convolution fitness value and greater than the convolution fitness value of the optimal convolution group is selected as the new optimal convolution group. Repeat until the convolution fitness of the optimal convolution group no longer changes, and output the optimal convolution group at this time as the convolution kernel of the convolution layer in the feature extraction model.

[0058] This embodiment uses artificial intelligence technologies such as deep learning to assist Chinese herbal medicine quality inspectors in conducting intelligent detection of palisade cells of Cuscuta australis and improve the accuracy and efficiency of screening work, and has significant application prospects. By combining deep learning technology with knowledge of Chinese herbal medicine identification and detection, a digital microscopic imaging-based intelligent detection technology for palisade cells of Cuscuta australis is realized, which further improves the accuracy and efficiency of Chinese herbal medicine identification and detection, effectively solves the defects in the detection process of palisade cells of Cuscuta australis, and promotes the application and development of a new generation of artificial intelligence technology in intelligent detection and auxiliary identification of Chinese herbal medicines. Through an efficient automated training process, the need for manual intervention is reduced, the level of intelligent detection is improved, and the needs of modern biological research and clinical applications are met.

[0059] In the embodiment of the present invention, thousands of high-quality, high-resolution digital microscopic images of palisade cells of Cuscuta australis obtained in cooperation with the Institute for the Authentication of Traditional Chinese and Ethnomedicines were used to train the neural network model. The labeling work of the present invention was completed by a number of professional Chinese medicine material inspectors, and a total of more than 1,200 digital microscopic images of palisade cells of Cuscuta australis were labeled.

[0060] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the embodiments of the present application.

[0061] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Similarly, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal device. In the absence of further restrictions, the elements defined by the sentence "comprise one..." do not exclude the presence of other identical elements in the process, method, article or terminal device including the elements.

[0062] The above is a detailed introduction to the training method and device for the intelligent detection model of palisade cells of southern dodder provided in the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for general technical personnel in this field, according to the idea of ​​the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A method for training an intelligent detection model of palisade cells of Cuscuta australis, characterized in that: include: S1. Collect digital microscopic images of palisade cells of Cuscuta australis, and each digital microscopic image has feature annotations; based on Feature annotation performs random aggregation on digital microscopic images to obtain A microscopic imaging clusters, and then splits the microscopic imaging clusters into clusters to obtain training clusters and optimization clusters; S2. Constructing the initial intelligent detection model of palisade cells of Cuscuta australis based on deep learning technology; S3, using the training cluster to cross-train the initial southern dodder palisade cell intelligent detection model to obtain the most suitable southern dodder palisade cell intelligent detection model; S4. Use the optimization cluster to iteratively optimize the most suitable southern dodder palisade cell intelligent detection model to obtain the target southern dodder palisade cell intelligent detection model.

2. The method for training an intelligent detection model of palisade cells of Cuscuta australis according to claim 1, characterized in that: The feature annotation includes: feature classification annotation and target positioning annotation; the target positioning annotation includes: marking position coordinates on digital microscopic imaging; the feature classification annotation includes: marking image features of palisade cells of southern dodder on digital microscopic imaging.

3. The method for training an intelligent detection model of palisade cells of Cuscuta australis according to claim 2, characterized in that: The method of randomly aggregating digital microscopic imaging comprises: Based on feature annotation, a microscopic feature vector is constructed. The similarity between different microscopic feature vectors is calculated using a similarity calculation formula. Based on the similarity, a similarity matrix M is constructed. ij represents the element in the i-th row and j-th column of the similarity matrix, then M ij represents the similarity between the i-th microscopic feature vector and the j-th microscopic feature vector; for the i-th row of the similarity matrix, the indexes of the columns whose element values ​​are greater than the preset similarity threshold constitute the neighboring set of the i-th microscopic feature vector; the fitness of each microscopic feature vector is calculated based on the similarity matrix and the neighboring set, and the formula for fitness calculation is: Where D(i) represents the local fitness of the i-th microscopic feature vector, N represents the neighboring set, Sim represents the size of the neighboring set, and σ represents the adjustment parameter. Cluster analysis is performed through local fitness to classify similar digital microscopic images into one category, and A microscopic imaging clusters are obtained, where A is an integer greater than zero.

4. The method for training an intelligent detection model of palisade cells of Cuscuta australis according to claim 3, characterized in that: The construction method of the initial Cuscuta australis palisade cell intelligent detection model includes: The initialization feature extraction model is a neural network model. The network layer structure includes: input layer, convolution layer, pooling layer, fully connected layer and output layer. The Sigmoid function is used as the activation function. The input layer uses the training set as input, and the output layer uses the feature cell map as output. The feature recognition model is initialized as the Faster R-CNN model, and the region proposal network is defined to use the Anchor mechanism as the candidate box generation mechanism. The classifier is the SoftMax classifier, the regressor is the Lasso regressor, the feature cell map is used as the input of the feature recognition model, and the target recognition result is used as the output of the feature recognition model.

5. The method for training an intelligent detection model of palisade cells of Cuscuta australis according to claim 4, characterized in that: The method of cross-training the initial Cuscuta australis palisade cell intelligent detection model comprises: Based on the total number of clusters K in the training clusters, K model trainings are performed. Each time the model training takes one of the training clusters as a validation set, and the other training clusters are trained as training sets to obtain K initial southern dodder palisade cell intelligent detection models; the initial southern dodder palisade cell intelligent detection model includes a feature extraction model and a feature recognition model. For each model training, the digital microscopic imaging in the training set is used as the input of the feature extraction model, and the digital microscopic imaging in the training set is convolved through a convolution layer to obtain a feature map; the feature map is pooled through a pooling layer to obtain a low-dimensional feature map; the low-dimensional feature map is feature mapped through a fully connected layer to obtain a feature cell map; the feature recognition model is used to perform feature recognition on the feature cell map to obtain a target recognition result, and the target recognition result is confidence calculated to obtain a confidence parameter; the average value of the confidence parameters obtained in each training is calculated, and the initial southern dodder palisade cell intelligent detection model with the largest average value is used as the most suitable southern dodder palisade cell intelligent detection model.

6. The method for training an intelligent detection model of palisade cells of Cuscuta australis according to claim 5, characterized in that: The method of performing a convolution operation on the digital microscopic imaging in the training set includes: presetting a moving stride of the convolution kernel and a convolution matrix size, sliding the convolution kernel of the convolution layer in the digital microscopic imaging with the moving stride, and multiplying and summing the pixel elements of the selected digital microscopic imaging area having the same size as the convolution matrix with the convolution kernel after each sliding, so as to extract local features of the digital microscopic imaging; The method of performing a pooling operation on the feature map includes: dynamically adjusting the pooling window based on the size of the acquired feature map and the size of the digital microscopic imaging, and the length calculation formula of the pooling window is: Among them, HEI represents the length of the pooling window, H in Represents the length of digital microscopy imaging, H out Represents the length of the feature map, Floor() represents the floor function; the width of the pooling window is calculated as: Among them, WID represents the width of the pooling window, W in Represents the width of the digital microscope image, W out Represents the width of the feature map; select elements in the feature map according to the length and width of the pooling window. When the length of the pooling window is equal to the width, take the average value of the elements in the pooling window. When the length of the pooling window is not equal to the width, take the maximum value of the elements in the pooling window.

7. The method for training an intelligent detection model of palisade cells of Cuscuta australis according to claim 6, characterized in that: The method of performing feature recognition on the characteristic cell map includes: The region proposal network scans the feature cell map and generates candidate boxes of different scales and aspect ratios using the Anchor mechanism. The position and size of the bounding box are adjusted through the regressor. The object of each candidate box and the category to which the object belongs are identified through the classifier. The candidate boxes identified with the same category are spliced ​​to obtain a new candidate box. According to the size of the feature cell map and the size of the new candidate box, the area selected by the candidate box is converted into regional data, and the converted regional data is the target recognition result.

8. The method for training an intelligent detection model of palisade cells of Cuscuta australis according to claim 7, characterized in that: The formula for calculating the confidence of the target recognition result is: Where C(x) represents the confidence parameter of the x-th target recognition result, All represents the total number of categories of the recognized content in the target recognition result, and S represents the number of categories in the target recognition result that are consistent with the feature annotation content. represents the correlation between the x-th target recognition result and the e-th category in the x-th feature annotation, β e Represents the weight value of the e-th category in the target recognition result, x e represents the e-th category of the x-th target recognition result, Represents the e-th category of the x-th feature annotation.

9. The method for training an intelligent detection model of palisade cells of Cuscuta australis according to claim 8, characterized in that: The method of iteratively optimizing the optimal palisade cell intelligent detection model of Cuscuta australis comprises: A genetic algorithm is used to perform convolution optimization on the convolution layer in the feature extraction model to obtain the optimal convolution group. The optimal convolution group is used as the convolution kernel of the convolution layer in the feature extraction model to obtain the target feature extraction model. The digital microscopic imaging in the optimized cluster is used as the input of the target feature extraction model for feature extraction to obtain feature cell maps corresponding to different optimized clusters. The cell map features corresponding to the same optimized cluster constitute a feature map cluster. The feature map cluster is used as the input of the feature recognition model for feature recognition to obtain the target recognition result. The target recognition result is trained by a comprehensive loss function. The non-maximum suppression algorithm is used to suppress the repeated results of different candidate boxes to obtain the target feature recognition model.

10. The method for training an intelligent detection model of palisade cells of Cuscuta australis according to claim 9, characterized in that: The method of performing convolution optimization on the convolution layer in the feature extraction model includes: For the convolution layer in the feature extraction model, P initial convolution kernel groups are preset, each convolution kernel group has E convolution kernels, the optimal convolution group is initialized to be empty, the weight parameters in each convolution kernel are initialized, and the weight parameters are randomly taken based on the normal distribution L(a, b); based on the convolution kernel in the convolution layer, the digital microscopic imaging in the optimized cluster is used as the input of the feature extraction model for forward propagation, the input of the previous convolution layer is used as the output of the next convolution layer, the cross entropy function is used as the loss function of each convolution layer, the convolution kernel group is traversed, each convolution kernel group is used as the convolution kernel in the convolution layer, and the inverse of the loss function value is used as the convolution fitness of each convolution kernel group; the selection probability of each convolution kernel group is calculated based on the convolution fitness, and the calculation formula of the selection probability is: Among them, Value(c) represents the selection probability of the cth convolution kernel group, d represents the index of the convolution kernel group, Fit(d) represents the convolution fitness of the dth convolution kernel group, and Fit(c) represents the convolution fitness of the cth convolution kernel group; iterative selection is performed based on the selection probability, and F convolution kernel groups are selected as the parent convolution kernel groups each time. The parent convolution kernel groups are combined in pairs using the permutation and combination method to obtain the parent convolution pairs. For each pair of parent convolution pairs, the convolution kernel with the largest modulus value in the parent convolution kernel group is selected as the crossover node for single-point crossover to obtain the child convolution kernel group; the increment value is preset, the convolution kernels in the child convolution kernel group are traversed, and a random selection algorithm is used. The weight parameters in the convolution kernel are selectively mutated, and each weight parameter that is selectively mutated is added to the incremental value to obtain a mutated convolution kernel group; the mutated convolution kernel group is used as the convolution kernel in the convolution layer, and the convolution fitness of the mutated convolution kernel group is calculated. When the optimal convolution group is empty, the mutated convolution kernel group with the largest convolution fitness value is selected as the new optimal convolution group. When the optimal convolution group is not empty, the mutated convolution kernel group with the largest convolution fitness value that is greater than the convolution fitness value of the optimal convolution group is selected as the new optimal convolution group. Repeat until the convolution fitness of the optimal convolution group no longer changes, and output the optimal convolution group at this time as the convolution kernel of the convolution layer in the feature extraction model.

Images