Microscopic identification method for Miao medicine compound eight-ingredient Huiyang powder based on improved YOLOv10

Abstract

The invention discloses an improved YOLOv10-based Miao medicine compound eight-ingredient reyang powder microscopic identification method, which is characterized in that a YOLOv10 network is improved based on a Triplet AtC2f module and a GFPN module, sufficient recall rate and stable high-IoU target anchor frame selection are ensured under the condition of microscopic category complex scale change, and meanwhile, the overall reasoning speed of a model is ensured. The method comprises the following steps: collecting and marking a Miao medicine compound eight-ingredient Reyang powder microscopic identification data set, sampling and marking each sample category in the data set, and dividing the data set into a training set and a test set; the YOLOv10 basic model is improved on the basis of a TripletAtC2f module and a GFPN module, and an improved YOLOv10 Miao medicine compound eight-ingredient reyang powder microscopic identification model is obtained; inputting the marked data in the training set into an improved YOLOv10 Miao medicine Huangyang Dan microscopic identification model for model training; and detecting data in the test set by using an optimal weight file of the improved YOLOv10 Miao medicine Huangyang Dan microscopic identification model to obtain anchor frames and confidence coefficients of different target categories in a microscopic image.

Description

Technical Field

[0001] The embodiments of the present application relate to the technical field of target detection based on a specific computing system, and more particularly to a method for microscopic identification of the Miao medicine compound Bawei Huanyang San based on improved YOLOv10. Background Art

[0002] Traditional Chinese medicine is a general term for Chinese medicine of all ethnic groups, including Han and ethnic minority medicine. Traditional Chinese medicine is a traditional Chinese medicine that is unique to traditional Chinese medicine. As traditional Chinese medicine is increasingly recognized around the world, the demand for Chinese medicinal materials has also increased dramatically. However, in order to pursue huge profits, some businesses do not hesitate to use various methods to forge, reduce costs, and seek huge profits. The quality of Chinese medicinal materials is worrying.

[0003] It is difficult for ordinary consumers and non-professionals to judge the quality and authenticity of Chinese herbal medicines with the naked eye, which also provides opportunities for illegal merchants. The quality and authenticity of Chinese herbal medicines are directly related to whether the medicines are safe. However, adulterated Chinese herbal medicines may also cause great harm and even lead to the death of patients. Therefore, it is particularly important to distinguish the authenticity of Chinese herbal medicines and ensure the quality of Chinese herbal medicines in order to combat and regulate the chaotic herbal medicine market.

[0004] Traditional methods for identifying Chinese herbal medicines are mainly original plant identification, trait identification, microscopic identification and physical and chemical identification. Original plant identification is only applicable in scientific research, and it is generally difficult to obtain original plants when conducting herbal medicine identification. Trait identification is mainly to identify the size, shape, taste and color of herbal medicines through vision, touch, smell or using water and fire to identify macroscopic characteristics. This method is simple and intuitive, but it is greatly affected by human subjective factors and requires the identifier to have rich professional knowledge and work experience.

[0005] Therefore, it is particularly important to establish a Chinese medicine identification scheme that combines deep learning and microscopic identification by establishing a database of Chinese medicine microscopic features. Deep learning has a strong sense of self-learning, as well as the ability to autonomously extract feature information and process data on a large scale. It can greatly reduce the threshold for Chinese medicine identification. Compared with identification based on human experience, it is more efficient and objective, with higher accuracy, and can effectively expand the application scenarios of microscopic identification of traditional Chinese medicinal material powder. At the same time, in the process of deep learning Chinese medicine identification, it also faces the factors of large changes in the scale of the detection target and the number of targets in the microscopic environment, which makes it difficult for the deep learning microscopic identification model to alleviate this type of interference. Therefore, there is an urgent need for a microscopic identification detection method to effectively alleviate sample interference and further improve detection accuracy. Summary of the invention

[0006] In order to overcome the above technical defects, the present application embodiment provides a microscopic identification method of Miao medicine compound Bawei Huanyang Powder based on improved YOLOv10, and its technical solution is as follows:

[0007] S1. Collect and annotate the microscopic identification dataset of Miao medicine compound Bawei Huanyang Powder, sample and annotate each sample category in the dataset, and divide it into training set and test set;

[0008] S2. The YOLOv10 basic model was improved based on the TripletAt_C2f module and the GFPN module to obtain the improved YOLOv10 microscopic identification model of the Miao medicine compound Bawei Huanyangsan;

[0009] S3, input the labeled data in the training set into the improved YOLOv10 Miao medicine Huangyangdan microscopic identification model for model training;

[0010] S4. Use the optimal weight file of the Miao medicine Huangyangdan microscopic identification model based on improved YOLOv10 to detect the data in the test set and obtain the anchor boxes and confidence levels of different target categories in the microscopic images.

[0011] Specifically, S2 improves the YOLOv10 basic model based on the TripletAt_C2f module, including:

[0012] S2.1. Replace the C2f module of the Backbone layer in the YOLOv10 network with the TriAT_C2f module. The TriAT_C2f module mainly receives the feature maps extracted from the corresponding backbone networks at each level. It increases the number of feature channels through two layers of convolution operations on the input feature maps and divides them into two sub-feature maps according to the number of channels. One sub-feature map is subjected to feature extraction through two layers of convolution operations and backed up as a feature map. The other sub-feature map enters the Bottleneck_TripletAt module to perform the triple attention Triplet Attention operation.

[0013] The Bottleneck_TripletAt module is the core module for implementing the triple attention operation in the TriAT_C2f module. The definition of the Bottleneck_TripletAt module includes two layers of convolution. The TripletAttention operation includes three branch operations: branch a is responsible for extracting features directly through residual transformation, branch b is responsible for extracting features through residual transformation after rotating the feature map along the feature width and channel dimensions, and branch c is responsible for extracting features through residual transformation after rotating the feature map along the feature height and channel dimensions.

[0014] Specifically, S2 improves the YOLOv10 basic model based on the GFPN module, including:

[0015] S2.2, the entire multi-scale feature aggregation and information interaction Neck layer of the YOLOv10 network is replaced by the GFPN module. The GFPN module mainly includes a CSPStage module that receives feature maps of different levels input by Backbone. The CSPStage module is used to perform the first-layer bottom-up re-parameter fusion operation and the second-layer top-down re-parameter fusion operation;

[0016] In the CSPStage module, N ResRepBlock modules are connected in series in the form of residual links, and the output of each module undergoes skip connection operations and re-parameter fusion operations.

[0017] Specifically, the calculation formula for the Triplet Attention operation is:

[0018]

[0019] Specifically, the CSPStage module can be expressed as:

[0020] RRB=Concat(F input ,Conv 1*1 (Rep(F input )))

[0021] F CSPStage =Concat(Conv 1*1 (F concat ),Conv 1*1 (RRB(F concat ),...,RRB(F concat ) N )).

[0022] Specifically, the parameter settings for model training in S3 are: batch size is 8, learning rate is 0.001, epoch is 100, and adam optimizer is used. During the model training process, the training results of each epoch are saved in the form of a weight file, and the file with the smallest inference loss among all weight files is used as the optimal weight file output of the model.

[0023] Specifically, labeling each sample category in the data set includes:

[0024] The graphic annotation tool was used to annotate 19 different target categories, including Eucommia ulmoides cork cells, Eucommia ulmoides rubber fibers, Eucommia ulmoides stone cells, mulberry stone cells, mulberry perianth thin-walled cells, Epimedium upper epidermal cells, Epimedium non-glandular hairs, Epimedium lower epidermis and stomata, Epimedium fibers and calcium oxalate columnar crystals, Cistanche deserticola vessels, Cistanche deserticola epidermal cells, Euryale ferox starch granules, Euryale ferox starch complex granules, Cuscuta australis cotyledon cells containing fuzzy powder and oil droplets, Cuscuta australis palisade cells, Cuscuta australis epidermal cells, Coix seed starch granules, Allium sphaerocarpa endosperm cells, and Allium sphaerocarpa epidermal cells.

[0025] The embodiments of the present application adopt the above technical solution to achieve the following technical effects:

[0026] In order to solve the problems of low recall rate and poor robustness of sample identification in multi-scale and multi-target microscopic environments, this application improves the YOLOv10 network based on the TripletAt_C2f module and the GFPN module: (1) The TripletAt_C2f module is used to replace the C2f module of the Backbone layer in the YOLOv10 network, which not only enhances the information interaction ability between layer-by-layer features, but also improves the context representation ability of the model to extract features. (2) The GFPN module is used to replace the main part of the Neck layer, and different structures are designed for the training stage and the reasoning stage based on the heavy parameter principle to ensure that the model has a larger amount of parameters for sufficient feature information mining in the training stage, and the speed of the reasoning process is ensured by the lightweight module in the reasoning stage. In addition, the overall GFPN module architecture can break through the limitation of the traditional FPN one-way propagation, take advantage of the two-way information propagation, effectively integrate multi-scale features, and alleviate the interference caused by multi-scale and multi-number target categories in the microscopic identification process. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The accompanying drawings exemplarily illustrate the embodiments and constitute a part of the specification, and together with the text description of the specification, are used to explain the exemplary implementation of the embodiments. The embodiments shown are for illustrative purposes only and do not limit the scope of the claims. In all drawings, the same reference numerals refer to similar but not necessarily identical elements.

[0028] Figure 1 A flowchart of the microscopic identification method of the Miao medicine compound Bawei Huanyang Powder based on improved YOLOv10 in this application;

[0029] Figure 2 This is a structural diagram of the microscopic identification model of the Miao medicine compound Bawei Huanyangsan that improves YOLOv10 in this application;

[0030] Figure 3 This is a schematic diagram of the structure of the TriAT_C2f module in this application;

[0031] Figure 4This is a schematic diagram of the structure of the GFPN module in this application;

[0032] Figure 5 This is a schematic diagram of the detection results of the microscopic identification method of the Miao medicine compound Bawei Huanyang Powder based on the improved YOLOv10 in this application. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical solutions and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0034] It should be noted that the descriptions involving "first", "second", etc. in the embodiments of the present application are only for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In addition, the technical solutions between the various embodiments can be combined with each other, but they must be based on the ability of ordinary technicians in the field to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by this application.

[0035] In the description of the present application, it should be understood that the numerical labels before the steps do not indicate the order in which the steps are executed, but are only used to facilitate the description of the present application and to distinguish each step, and therefore should not be understood as a limitation on the present application.

[0036] First, the following terms are explained:

[0037] Backbone layer: As the initialization encoding part of the entire model that receives image data, it lays a good information foundation for further in-depth processing of subsequent features with its stable information encoding capabilities.

[0038] C2f module: This module is a feature fusion module in the traditional YOLO series. It mainly uses residual connections to further deepen the extraction of original features while retaining the original feature information.

[0039] TriAT_C2f module: This module is one of the innovative modules in this application. Based on the residual connection of the traditional C2f module, the information interaction between data channels is enhanced through triple attention operation.

[0040] Bottleneck_TripletAt module: As the core module for implementing triple attention operation in the TriAT_C2f module, this module achieves effective fusion and interaction of feature information by exchanging the same input features in different ways in terms of channel, width, and height, encoding and merging them.

[0041] Neck layer: As the part of the entire model that receives the output of the Backbone layer, the role of the Neck layer is to extract more usable feature information from the initial encoded data input by Backbone through an effective hierarchical structure design.

[0042] GFPN module: This module breaks through the problem of limited two-way information transmission in traditional FPN modules. It adopts a two-way information interaction structure and can use CSPStage to more effectively integrate information at all levels and enrich feature information.

[0043] CSPStage module: As the main component of the GFPN module, this module can receive features at different levels on demand and is the basic module for the GFPN module to realize information interaction at various levels.

[0044] ResRepBlock module: As an important module for re-parameter encoding in the CSPStage module, this module uses the Rep sub-module to re-parameter encode the input features to improve the reliability of feature information fusion.

[0045] Rep module: As a key module for re-parameterization, this module is divided into two forms: training and reasoning. In the training form, in order to better perform precise encoding operations, the model performance is improved by adding convolution modules of different sizes; in the reasoning form, in order to ensure the reasoning speed of the actual operation scenario, only a single convolution module is used for reasoning. Since the training process has achieved high-precision fitting, using only a single convolution module in the reasoning process will not affect the reasoning accuracy, and can also improve the model reasoning speed.

[0046] Secondly, in order to facilitate those skilled in the art to understand the technical solutions provided in the embodiments of the present application, the relevant technologies are described below:

[0047] See also Figure 1 The microscopic identification method of the Miao medicine compound Bawei Huanyang Powder based on the improved YOLOv10 in this application includes:

[0048] S1. Collect and annotate the microscopic identification dataset of the Miao medicine compound Bawei Huanyang Powder, sample and annotate each sample category in the dataset, and divide it into training set and test set.

[0049] The preset microscopic identification scene includes the sampling data of the Miao medicine compound Bawei Huanyang Powder at 100 times and 400 times magnification. Use the graphic annotation tool to annotate. Specifically, 4735 microscopic identification data of the Miao medicine compound Bawei Huanyang Powder were collected, and graphic annotation tools were used to annotate 19 different target categories, including Eucommia ulmoides cork cells, Eucommia ulmoides rubber fibers, Eucommia ulmoides stone cells, mulberry stone cells, mulberry perianth thin-walled cells, Epimedium upper epidermal cells, Epimedium non-glandular hairs, Epimedium lower epidermis and stomata, Epimedium fibers and calcium oxalate columnar crystals, Cistanche deserticola vessels, Cistanche deserticola epidermal cells, Euryale ferox starch granules, Euryale ferox starch complex granules, Cuscuta australis cotyledon cells containing fuzzy powder and oil droplets, Cuscuta australis palisade cells, Cuscuta australis epidermal cells, Coix seed starch granules, Allium sphaerocarpa endosperm cells, and Allium sphaerocarpa epidermal cells, and the annotated data set was randomly divided into a ratio of training set:test set=9:1.

[0050] S2. The YOLOv10 basic model was improved based on the TripletAt_C2f module and the GFPN module to obtain the improved YOLOv10 microscopic identification model of the Miao medicine compound Bawei Huanyangsan.

[0051] The improved microscopic identification model of the Miao medicine compound Bawei Huanyang Powder is as follows Figure 2 As shown, the specific steps of model improvement include the following S2.1-S2.2:

[0052] S2.1. Replace the C2f module of the Backbone layer in the YOLOv10 network with the TriAT_C2f module.

[0053] like Figure 2 and Figure 3 As shown in Figure 2, the TriAT_C2f module mainly receives feature maps extracted from the corresponding backbone networks at each level, which are represented by F∈R C×H×W The TriAT_C2f module performs two layers of convolution on the input feature map, increases the number of feature channels to 2C, and divides it into two sub-feature maps F with C channels. a 、F b Among them, F a Further feature extraction will be performed through two layers of convolution operations and backed up as feature maps. b It will continue to enter the Bottleneck_TripletAt module to perform the next triple attention convolution operation, namely the TripletAttention operation.

[0054] The definition of the Bottleneck_TripletAt module includes two layers of convolution and the execution of the TripletAttention operation, which is used to extract features from three different dimensional combination modes to fully integrate the information of each dimension and improve the overall robustness of the model. The TripletAttention operation will be divided into three branches: branch a is responsible for extracting features directly through residual transformation, branch b is responsible for extracting features through residual transformation after rotating the feature map along the feature width and channel dimensions, and branch c is responsible for extracting features through residual transformation after rotating the feature map along the feature height and channel dimensions.

[0055] The three branches establish inter-dimensional dependencies for cross-dimensional features, effectively enhancing the feature representation effect and reducing the computational consumption of spatial features. At the same time, the features of each branch are subjected to the dot multiplication operation of the original feature map after the Z-Pool and convolution operations, realizing the fusion and strengthening of the relationship between features of different dimensions and completing the entire TripletAttention operation.

[0056] The Z-Pool operation can be expressed as: Z-Pool(X) = [Max Pool 0d (X), Avg Pool 0d (X)], where a represents the input feature of branch a or the rotated feature of branch b and c. MaxPool 0d 、AvgPool 0d They represent the maximum pooling operation and the average pooling operation respectively. The Z-Pool operation compresses the channel dimension to 2 dimensions, connects the maximum value feature and the average value feature, and further reduces the amount of calculation while ensuring the feature representation of rich features. The entire TripletAttention operation can be fully described as:

[0057]

[0058] In the formula, Represent the corresponding output characteristics of each branch in the above Triplet Attention operation sequence. i In , Permute1 refers to the exchange of channel dimension and height dimension, followed by Z-Pool operation along the height dimension to capture the interaction between channel dimension and width dimension. Similarly, Permute2 refers to the exchange of channel dimension and width dimension, followed by Z-Pool operation along the width dimension to capture the interaction between channel dimension and height dimension. Channel Pool and AvgPool represent channel pooling and average pooling operations on input features, respectively. Concat, BN and They refer to channel connection operation, batch normalization and Sigmoid operation respectively. ⊙ represents element multiplication operation.

[0059] S2.2. Replace the entire multi-scale feature aggregation and information interaction Neck layer of the YOLOv10 network with the GFPN module.

[0060] Specifically, the GFPN module mainly includes the CSPStage module. Figure 4 As shown, the GFPN module is defined to receive feature maps of different levels input by Backbone. The feature maps are represented as:

[0061] Where i represents the level of input features, and j represents the level of fused features.

[0062] first, The two feature maps are input to the CSPStage module for the first layer of heavy parameter fusion output feature The three features are input into CSPSatge together for the first layer of heavy parameter fusion output Features are output by first layer parameter fusion So far, the first layer of bottom-up re-parameter fusion is completed. However, in order to prevent the single information interaction problem brought by one-way fusion, this module will continue to perform the second layer of top-down re-parameter fusion operation. Input to the CSPStage module for the second layer of heavy parameter fusion output Input to the CSPStage module for the second layer of heavy parameter fusion output At this point, the entire GFPN module has completed the double-layer re-parameter feature fusion operation; among them, the CSPStage module is very important for the re-parameter fusion process between various features.

[0063] After completing the preliminary channel cascade on the input feature image combination, this module performs a dual-branch independent Conv1*1 convolution operation for the final residual jump connection operation and further heavy parameter fusion operation.

[0064] To ensure the stability of re-parameter fusion, the CSPStage module contains the ResRepBlock module, which connects N ResRepBlock modules in series in the form of residual links. The output of each module is subjected to skip connection operations and further re-parameter fusion operations. In the ResRepBlock module, the Rep module is the direct action module of the re-parameter fusion principle. The Rep module has two framework structures, which are used for training and reasoning processes respectively. Since the single size of the convolution kernel will cause the problem of poor fusion effect, the training framework of the Rep model will use Conv3*3 and Conv1*1 operations to increase the amount of training parameters to perform dual-branch feature extraction on the input features at multiple scales and add the output feature maps, and finally output them through the SiLU activation function. In order to ensure that the reasoning speed of the model is not affected by the Rep module training framework, the Rep reasoning framework only uses one Conv3*3 operation to complete the fast reasoning of the re-parameter module. Because the Conv1*1 operation in the Rep training framework matches the Conv3*3 operation in the inference framework, the Conv3*3 operation Padding value is adjusted to simulate the Conv3*3 convolution operation. The CSPStage module can be further described as:

[0065] RRB=Concat(F input ,Conv 1*1 (Rep(F input )))

[0066] F CSPStage =Concat(Conv 1*1 (F concat ),Conv 1*1 (RRB(F concat ),...,RRB(F concat ) N )),

[0067] In the formula, F CSPStage Represents the output features of the CSPStage module. concat Represents the channel cascade feature of CSPStage input. input Represents the characteristics of the input RRB module.

[0068] S3. Input the labeled data in the training set into the improved YOLOv10 Miao medicine Huangyangdan microscopic identification model for model training.

[0069] The parameter settings during model training are: batchsize is 8, learning rate is 0.001, epoch is 100, and adam optimizer is used. Among them, batchsize is the number of images input to the module at one time; learning rate is a key parameter in the model iteration process, which is optimized and adjusted by adam optimizer; epoch is the number of complete training rounds. Finally, the training results of each epoch are saved in the form of weight files, and the file with the smallest inference loss among all weight files is output as the optimal weight file of the model.

[0070] S4. Use the optimal weight file of the Miao medicine Huangyangdan microscopic identification model based on improved YOLOv10 to detect the data in the test set and obtain the anchor boxes and confidence levels of different target categories in the microscopic images.

[0071] The anchor boxes and confidence scores of different target categories are the test results. The detection test specifically includes the following steps:

[0072] S4.1. Read the preset batchsize images in the test, and detect them through the model weight file to obtain the anchor box and confidence detection results of the target category in the image. Figure 5 shown.

[0073] S4.2. Re-read the images in the test set until all images in the test set have been detected.

[0074] Using the labeled dataset and setting the confidence threshold to 0.5, we can get the overall performance of the model.

[0075] Finally, in order to verify the improved model detection performance in this application, the microscopic identification method of the Miao medicine compound Bawei Huanyang Powder in this application was compared and verified with the comprehensive performance of each category of YOLOv5, YOLOv6, YOLOv7, YOLOv8, YOLOv9 and YOLOv10, and the performance index results are shown in the following table.

[0076] method Recall F1 score Average accuracy @50 Average accuracy @50-95 YOLOv5 0.846 0.830 0.916 0.763 YOLOv6 0.783 0.781 0.886 0.744 YOLOv7 0.906 0.880 0.930 0.796 YOLOv8 0.857 0.878 0.939 0.796 YOLOv9 0.866 0.857 0.926 0.787 YOLOv10 0.875 0.893 0.934 0.827 This application 0.914 0.910 0.950 0.848

[0077] From the performance indicators such as recall rate, F1 score and accuracy in the above table, we can intuitively demonstrate the ability and effect of the microscopic identification method of the Miao medicine compound Bawei Huanyang Powder in this application.

[0078] Obviously, those skilled in the art should understand that the modules or steps of the above-mentioned embodiments of the present application can be implemented by general-purpose computer devices, they can be concentrated on a single computer device, or distributed on a network composed of multiple computer devices, optionally, they can be implemented by executable program codes of computer devices, so that they can be stored in a storage device and executed by the computer device, and in some cases, the steps shown or described can be executed in a different order from that herein, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. In this way, the embodiments of the present application are not limited to any specific combination of hardware and software.

[0079] It should be noted that the above are only preferred embodiments of the present application, and the patent protection scope of the present application is not limited thereto. Any equivalent structure or equivalent process transformation made using the contents of the specification and drawings of the present application, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A microscopic identification method of Miao medicine compound Bawei Huanyang Powder based on improved YOLOv10, characterized in that: include: S1. Collect and annotate the microscopic identification dataset of Miao medicine compound Bawei Huanyang Powder, sample and annotate each sample category in the dataset, and divide it into training set and test set; S2. The YOLOv10 basic model was improved based on the TripletAt_C2f module and the GFPN module to obtain the improved YOLOv10 microscopic identification model of the Miao medicine compound Bawei Huanyangsan; S3, input the labeled data in the training set into the improved YOLOv10 Miao medicine Huangyangdan microscopic identification model for model training; S4. Use the optimal weight file of the Miao medicine Huangyangdan microscopic identification model based on improved YOLOv10 to detect the data in the test set and obtain the anchor boxes and confidence levels of different target categories in the microscopic images.

2. The microscopic identification method of the Miao medicine compound Bawei Huanyang Powder according to claim 1 is characterized in that: In S2, the YOLOv10 basic model is improved based on the TripletAt_C2f module, including: S2.

1. Replace the C2f module of the Backbone layer in the YOLOv10 network with the TriAT_C2f module. The TriAT_C2f module mainly receives the feature maps extracted from the corresponding backbone networks at each level. It increases the number of feature channels through two layers of convolution operations on the input feature maps and divides them into two sub-feature maps according to the number of channels. One sub-feature map is subjected to feature extraction through two layers of convolution operations and backed up as a feature map. The other sub-feature map enters the Bottleneck_TripletAt module to perform the triple attention Triplet Attention operation. The Bottleneck_TripletAt module is the core module for implementing the triple attention operation in the TriAT_C2f module. The definition of the Bottleneck_TripletAt module includes two layers of convolution. The TripletAttention operation includes three branch operations: branch a is responsible for extracting features directly through residual transformation, branch b is responsible for extracting features through residual transformation after rotating the feature map along the feature width and channel dimensions, and branch c is responsible for extracting features through residual transformation after rotating the feature map along the feature height and channel dimensions.

3. The microscopic identification method of the Miao medicine compound Bawei Huanyang Powder according to claim 1 is characterized in that: In S2, the YOLOv10 basic model is improved based on the GFPN module, including: S2.2, the entire multi-scale feature aggregation and information interaction Neck layer of the YOLOv10 network is replaced by the GFPN module. The GFPN module mainly includes a CSPStage module that receives feature maps of different levels input by Backbone. The CSPStage module is used to perform the first-layer bottom-up re-parameter fusion operation and the second-layer top-down re-parameter fusion operation; In the CSPStage module, N ResRepBlock modules are connected in series in the form of residual links, and the output of each module undergoes skip connection operations and heavy parameter fusion operations.

4. The microscopic identification method of the Miao medicine compound Bawei Huanyang Powder according to claim 2 is characterized in that: The calculation formula for the TripletAttention operation is:

5. The microscopic identification method of the Miao medicine compound Bawei Huanyang Powder according to claim 3 is characterized in that: The CSPStage module can be expressed as: RRB=Concat(F input ,Conv 1*1 (Rep(F input ))) F CSPStage =Concat(Conv 1*1 (F concat ),Conv 1*1 (RRB(F concat ),...,RRB(F concat ) N ))。 6. The microscopic identification method of the Miao medicine compound Bawei Huanyang Powder according to claim 1 is characterized in that: The parameter settings for model training in S3 are: batchsize is 8, learning rate is 0.001, epoch is 100, and the adam optimizer is used. During the model training process, the training results of each epoch are saved in the form of a weight file, and the file with the smallest inference loss among all weight files is output as the optimal weight file of the model.

7. The method for microscopic identification of the Miao medicine compound Bawei Huanyang Powder according to claim 1, characterized in that: Labeling each sample category in the data set specifically includes: The graphic annotation tool was used to annotate 19 different target categories, including Eucommia ulmoides cork cells, Eucommia ulmoides rubber fibers, Eucommia ulmoides stone cells, mulberry stone cells, mulberry perianth thin-walled cells, Epimedium upper epidermal cells, Epimedium non-glandular hairs, Epimedium lower epidermis and stomata, Epimedium fibers and calcium oxalate columnar crystals, Cistanche deserticola vessels, Cistanche deserticola epidermal cells, Euryale ferox starch granules, Euryale ferox starch complex granules, Cuscuta australis cotyledon cells containing fuzzy powder and oil droplets, Cuscuta australis palisade cells, Cuscuta australis epidermal cells, Coix seed starch granules, Allium sphaerocarpa endosperm cells, and Allium sphaerocarpa epidermal cells.

Images