A traditional Chinese medicine microscopic identification method and system based on a multi-dimensional channel attention mechanism

Abstract

The application discloses a traditional Chinese medicine microscopic identification method and system based on a multi-dimensional channel attention mechanism, and comprises the following steps: combining the characteristics of incomplete cell structure, image stereoscopicness and uneven feature distribution of a traditional Chinese medicine microscopic image in an Input layer, obtaining a traditional Chinese medicine microscopic image data enhancement model, and enhancing traditional Chinese medicine microscopic image data by using the enhancement model; fusing shallow feature information in a Backbone layer, fusing deep feature information before a Prediction layer after a Neck layer, and obtaining a multi-dimensional channel attention microscopic feature extraction model; obtaining implicit effective auxiliary information from different channels from the shallow layer to the deep layer by using the extraction model; generating a prediction frame by using the Prediction layer to predict the features of the traditional Chinese medicine microscopic image by using the implicit effective auxiliary information, obtaining the features of the cells of the traditional Chinese medicine microscopic image, completing the identification of the cells of the traditional Chinese medicine microscopic image, and thereby improving the prediction accuracy of the traditional Chinese medicine microscopic image.

Description

Technical Field

[0001] This invention relates to the field of microscopic image information processing technology, specifically to a method and system for microscopic recognition of traditional Chinese medicine based on a multidimensional channel attention mechanism. Background Technology

[0002] With the booming development of the Chinese medicinal materials market, the business of testing and identification of Chinese medicinal materials is also thriving. Currently, wild medicinal materials are scarce, while cultivated materials are abundant, leading to cell variation. Identification based on experience and microscopic images results in low recognition rates and low specificity. Furthermore, due to the lack of a complete cell data reference library, traditional microscopic images and other data are scattered, incomplete, and slowly updated, with many subjective factors in the description of target substances.

[0003] Traditional methods for identifying Chinese medicinal materials mainly include identification of the origin, morphological identification, microscopic identification, and physicochemical identification. Among these, microscopic identification is based on the principle that different medicinal materials have different microscopic characteristics, using a microscope to observe the tissue structure and powder characteristics of the medicinal materials to distinguish between genuine and counterfeit products. Currently, due to the decreasing cost of microscopes year by year, microscopic identification has become the main identification method for pharmaceutical companies, hospitals, and drug regulatory agencies.

[0004] In recent years, thanks to the development of artificial neural networks, numerous image recognition methods based on deep learning have emerged. Image recognition (also known as object detection) methods have been widely used to accurately detect and locate objects in images by building neural network models to learn from data. Deep learning-based object detection algorithms are mainly divided into two categories: one is the two-stage detection algorithm based on region proposal; the other is the one-stage detection algorithm based on regression. Compared with two-stage detection algorithms, one-stage detection algorithms are faster, and representative algorithms include SSD, RetinaNet, and the YOLO series.

[0005] Deep learning-based object detection algorithms are widely used across various industries, especially the YOLO series of One-Stage detection algorithms. Among the YOLO series, the YOLOv5 network is widely used due to its ease of deployment in production using the PyTorch framework, and its advantages such as user-friendliness, ease of training, fast training speed, and high accuracy. However, because the YOLOv5 network performs poorly in detecting small targets, the following problems exist when predicting targets in microscopic images of traditional Chinese medicine:

[0006] (1) Due to problems such as incomplete structure, three-dimensionality, and uneven feature distribution in microscopic images of traditional Chinese medicine, the prediction accuracy is low.

[0007] (2) When identifying microscopic images of traditional Chinese medicine, the accuracy of prediction is further reduced because important data is lost as the number of layers of the convolutional neural network increases, and there is an information gap between low-dimensional feature maps and high-dimensional feature maps. Summary of the Invention

[0008] To overcome the shortcomings of existing technologies, this invention provides a method and system for microscopic identification of traditional Chinese medicine based on a multidimensional channel attention mechanism. This method addresses the technical problem of low accuracy in predicting microscopic images of traditional Chinese medicine using existing microscopic image recognition technologies, thereby improving the accuracy of microscopic image prediction of traditional Chinese medicine.

[0009] To solve the above problems, the technical solution adopted by the present invention is as follows:

[0010] A method for microscopic identification of traditional Chinese medicine based on a multidimensional channel attention mechanism includes the following steps:

[0011] By combining the characteristics of incomplete cell structure, three-dimensional image, and uneven feature distribution of traditional Chinese medicine microscopic images in the Input layer, a traditional Chinese medicine microscopic image data enhancement model is obtained, and the traditional Chinese medicine microscopic image data enhancement model is used to enhance the traditional Chinese medicine microscopic image data.

[0012] Shallow feature information is fused in the Backbone layer, and deep feature information is fused after the Neck layer and before the Prediction layer to obtain a microscopic feature extraction model with multi-channel attention.

[0013] The microscopic feature extraction model of the multidimensional channel attention is used to obtain implicit and effective auxiliary information from different channels from shallow to deep layers.

[0014] Using the implicit and effective auxiliary information, a prediction box is generated through the Prediction layer, and the prediction box is used to predict the features of the microscopic image of traditional Chinese medicine to obtain the features of the cells in the microscopic image of traditional Chinese medicine, thus completing the identification of the cells in the microscopic image of traditional Chinese medicine.

[0015] In a preferred embodiment of the present invention, when enhancing the data using the traditional Chinese medicine microscopic image data enhancement model, the following steps are included:

[0016] Randomly select n images from the original training dataset, and perform horizontal average segmentation on the n images to obtain several segmented images;

[0017] From the aforementioned segmented images, n images are randomly selected and horizontally stitched together to obtain several stitched images.

[0018] A new training dataset is constructed using the stitched images and the images in the original training dataset.

[0019] In a preferred embodiment of the present invention, constructing a new training dataset includes:

[0020] The stitched images and the images in the original dataset are then subjected to mirroring, translation, and rotation processing.

[0021] In a preferred embodiment of the present invention, when fusing feature information from shallow layers in the Backbone layer, the following is included:

[0022] By incorporating the SENet attention mechanism into the shallow layers of the network, a shallow channel attention mechanism SEAtt is obtained.

[0023] The first feature image, after being enhanced by the traditional Chinese medicine microscopic image data enhancement model, is processed by the feature processing module and transformed into a second feature image. The second feature image is then used as the input to the SEAt module.

[0024] The SEAt module aggregates the second feature image at different granularities to form image features and extracts them, thus completing the fusion of shallow feature information.

[0025] In a preferred embodiment of the present invention, the process of aggregating and extracting image features at different granularities includes:

[0026] The second feature image is transformed by a series of convolution operations to obtain the image features. ;

[0027] The image features are processed using the Squeeze operation in the SENet module. Compression yields a residual channel statistic;

[0028] Based on the residual channel statistics, and by predicting the importance of each channel using the Excitation operation in the SENet module, the nonlinear relationship between different channels is obtained.

[0029] Based on the image features The nonlinear relationship between the different channels is output through the Scale operation in the SENet module;

[0030] The Excitation operation employs a Sigmoid-based gating mechanism.

[0031] In a preferred embodiment of the present invention, when fusing deep feature information after the Neck layer and before the Prediction layer, the following is included:

[0032] By introducing the multidimensional deep channel attention mechanism MCATt into the deep layer of the network, the ECA attention mechanism is reconstructed for semantic and localization information of different dimensions and levels, providing a foundation for the accurate fusion of low, medium and high dimension feature maps at the deep layer of the network.

[0033] By combining the shallow channel attention mechanism SEAtt and the multidimensional deep channel attention mechanism MCAT, a microscopic feature extraction model for the multidimensional channel attention is obtained.

[0034] The deep layers of the network include low, medium, and high dimensions.

[0035] As a preferred embodiment of the present invention, when providing a foundation for the accurate fusion of low, medium, and high-dimensional feature maps in the deep layers of the network, it includes:

[0036] Input low-dimensional features Mid-dimensional characteristics and high-dimensional features After a series of convolution operations, they are transformed into features. ,feature and features ;

[0037] Perform a Squeeze operation to compress the features along the spatial dimension using global average pooling. The features and the aforementioned features After compression, one-dimensional real numbers are obtained respectively. One-dimensional real numbers and one-dimensional real numbers ;

[0038] According to the real number The real number and the real number The ECA module uses a one-dimensional convolution with a kernel size of k to perform cross-channel information interaction, and obtains low-dimensional weights, medium-dimensional weights and high-dimensional weights respectively.

[0039] The low-dimensional weights, medium-dimensional weights, and high-dimensional weights are combined with the original feature maps and output through the Scale operation in the ECA module to obtain low-dimensional residual features, medium-dimensional residual features, and high-dimensional residual features, respectively, providing a foundation for the accurate fusion of low, medium, and high-dimensional feature maps.

[0040] The Neck layer and the Prediction layer each have low-dimensional features. Mid-dimensional characteristics and high-dimensional features The feature output has three different dimensions.

[0041] In a preferred embodiment of the present invention, when generating a prediction box through a prediction layer and using the prediction box to predict the features of a microscopic image of traditional Chinese medicine, the process includes:

[0042] By using a microscopic identification model for traditional Chinese medicine, features from different receptive fields are fused in three different dimensions (low, medium, and high) to generate feature maps at three different target scales.

[0043] The feature maps of the three different target scales are used to predict targets in microscopic images of traditional Chinese medicine.

[0044] In a preferred embodiment of the present invention, predicting targets in microscopic images of traditional Chinese medicine includes:

[0045] The input microscopic image of Chinese medicine is divided into grids to obtain the grids where the center of the target exists, and the target is predicted using the grids.

[0046] The microscopic images of traditional Chinese medicine are predicted during the training phase using a loss consisting of three parts: target loss, category loss, and confidence loss.

[0047] Non-maximum suppression is used to remove redundant prediction boxes, select high-quality detection results, and obtain the characteristics of cells in the microscopic images of traditional Chinese medicine.

[0048] A microscopic recognition system for traditional Chinese medicine based on a multidimensional channel attention mechanism, comprising:

[0049] Enhancement unit: used to combine the characteristics of incomplete cell structure, three-dimensional image, and uneven feature distribution of traditional Chinese medicine microscopic images in the Input layer to obtain a traditional Chinese medicine microscopic image data enhancement model, and to enhance the traditional Chinese medicine microscopic image data using the traditional Chinese medicine microscopic image data enhancement model;

[0050] Auxiliary information acquisition unit: used to fuse shallow feature information in the Backbone layer and deep feature information after the Neck layer and before the Prediction layer to obtain a multi-dimensional channel attention microscopic feature extraction model; through the multi-dimensional channel attention microscopic feature extraction model, implicit effective auxiliary information is obtained from different channels from shallow to deep layers;

[0051] Prediction unit: Used to generate a prediction box through the Prediction layer using the implicit effective auxiliary information, and use the prediction box to predict the features of the microscopic image of Chinese medicine, thereby obtaining the features of the cells in the microscopic image of Chinese medicine and completing the identification of the cells in the microscopic image of Chinese medicine.

[0052] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0053] (1) In view of the problems of incomplete structure, three-dimensionality of images and uneven distribution of features in traditional Chinese medicine microscopic images, this invention proposes a data enhancement model for traditional Chinese medicine microscopic images, which can expand and balance the existing traditional Chinese medicine microscopic image dataset, and provide a data foundation for the effective prediction of traditional Chinese medicine microscopic images;

[0054] (2) This invention addresses the problem of important data loss in the identification of Chinese medicinal materials using microscopic image recognition technology as the number of layers in the convolutional neural network increases, and considers the information gap between low-dimensional and high-dimensional feature maps. It constructs a multi-channel attention-based microscopic feature extraction model to obtain implicit and effective auxiliary information from different channels in shallow and deep layers (including low, medium, and high dimensions), thus providing information for low, medium, and high-dimensional feature maps.

[0055] Accurate fusion of feature maps of the same dimension provides a foundation for enhancing prediction accuracy on feature maps of different sizes;

[0056] (3) This invention constructed a small dataset of 12,000 data points for two kinds of traditional Chinese medicine and conducted a series of comparative experiments. The results show that the MCA-CHMR method of this invention has the best performance compared with methods such as YOLOv5, SSD, RetinaNet and FASTER RCNN.

[0057] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0058] Figure 1 - This is a flowchart illustrating the steps of a microscopic identification method for traditional Chinese medicine based on a multidimensional channel attention mechanism according to an embodiment of the present invention;

[0059] Figure 2 - is a schematic diagram of the network structure of MCA-CHMR according to an embodiment of the present invention;

[0060] Figure 3 - This is a diagram illustrating the processing steps of the traditional Chinese medicine microscopic image data enhancement model according to an embodiment of the present invention;

[0061] Figure 4 - This is a comparative experimental diagram of Scutellaria baicalensis Vessels and Scutellaria baicalensis Corkcells in an embodiment of the present invention;

[0062] Figure 5 - This is a first comparative experimental diagram of Scutellaria baicalensis Corkcells and Magnolia officinalis Stonecells in an embodiment of the present invention;

[0063] Figure 6 - This is a second comparative experimental diagram of Scutellaria baicalensis Corkcells and Magnolia officinalis Stonecells in an embodiment of the present invention;

[0064] Figure 7- is the AP curve of the comparative experiment of Magnolia officinalis Stonecells in the embodiment of the present invention;

[0065] Figure 8 - These are comparative experimental figures of Magnolia officinalis Fibres according to embodiments of the present invention. Detailed Implementation

[0066] The microscopic identification method for traditional Chinese medicine based on a multidimensional channel attention mechanism provided by this invention, such as... Figure 1 As shown, it includes the following steps:

[0067] Step S1: Combining the characteristics of incomplete cell structure, three-dimensional image, and uneven feature distribution in the input layer of traditional Chinese medicine microscopic images, a data enhancement model for traditional Chinese medicine microscopic images is obtained, and the data enhancement model for traditional Chinese medicine microscopic images is used to enhance the data of traditional Chinese medicine microscopic images.

[0068] Step S2: Fuse shallow feature information in the Backbone layer and fuse deep feature information after the Neck layer and before the Prediction layer to obtain a multi-channel attention microscopic feature extraction model.

[0069] Step S3: Extract implicit and effective auxiliary information from different channels from shallow to deep layers using a multi-channel attention microscopic feature extraction model;

[0070] Step S4: Using implicit and effective auxiliary information, a prediction box is generated through the Prediction layer, and the prediction box is used to predict the features of the microscopic image of Chinese medicine to obtain the features of the cells in the microscopic image of Chinese medicine, thus completing the identification of cells in the microscopic image of Chinese medicine.

[0071] Specifically, this invention proposes a microscopic identification method for traditional Chinese medicine based on a multidimensional channel attention mechanism (MCA-CHMR) on the basis of YOLOv5. For example... Figure 2 As shown, firstly, a Traditional Chinese Medicine (TCM) microscopic image data augmentation model (TCM-MIEM) is designed at the Input layer, taking into account the characteristics of incomplete cell structure, three-dimensionality, and uneven feature distribution in TCM microscopic images. This model can expand and balance existing TCM microscopic image datasets. Secondly, a multi-channel attention microscopic feature extraction model (SEAtt+MCAtt) is proposed, fusing shallow feature information at the Backbone layer and deep feature information after the Neck layer and before the Prediction layer. This model extracts implicit and effective auxiliary information from different channels (including low, medium, and high dimensions). Finally, the image features are predicted through the Prediction layer to generate bounding boxes and obtain the final result. The MCA-CHMR method proposed in this invention is also divided into four main modules: Input, Backbone, Neck, and Prediction, which are executed sequentially.

[0072] In step S1 above, when enhancing the data using a traditional Chinese medicine microscopic image data enhancement model, the following steps are included:

[0073] Randomly select n images from the original training dataset and perform horizontal average segmentation on the n images to obtain several segmented images;

[0074] From a number of segmented images, n images are randomly selected and horizontally stitched together to obtain a number of stitched images.

[0075] A new training dataset is constructed using several stitched images and images from the original training dataset.

[0076] Specifically, such as Figure 2 As shown, this invention proposes a Traditional Chinese Medicine Microscopic Image Data Enhancement Model (TCM-MIEM) in the Input stage of MCA-CHMR. The specific operation process of TCM-MIEM is as follows: Figure 3 As shown in the diagram, first, two images, a and b, are randomly selected from the dataset; then, these two images are horizontally segmented into an average size; finally, two more images are randomly selected from the segmented images and horizontally stitched together. The stitched image can be any combination other than a and b.

[0077] For example, if the number of expanded images is M, then the data size changes from the original N to N+M, i.e., i∈[1, N+M], X={X1,X2,...,X...} i ,...,X N+M}

[0078] Considering that it is impossible to capture an image for every real scene that the model of this invention will see in inference, in order to further expand the data scale, this invention, based on the traditional Chinese medicine microscopic image data enhancement model, also references YOLOv5 to introduce processing such as mirroring, translation and rotation.

[0079] In this invention, the TCM-MIEM method is used to augment the data, and the dataset is then input into the network as entirely new training data to complete the training. Overall, the TCM-MIEM method proposed in this invention can expand and balance existing datasets of microscopic images of traditional Chinese medicine, providing a data foundation for the effective prediction of microscopic images of traditional Chinese medicine.

[0080] Furthermore, when constructing a completely new training dataset, the following steps are included:

[0081] The stitched images and the images in the original dataset are then mirrored, translated, and rotated.

[0082] In step S2 above, when fusing shallow feature information in the Backbone layer, the following is included:

[0083] By incorporating the SENet attention mechanism into the shallow layers of the network, a shallow channel attention mechanism SEAtt is obtained.

[0084] The first feature image, which is output after being enhanced by the traditional Chinese medicine microscopic image data enhancement model, is processed by the feature processing module and transformed into a second feature image. The second feature image is then used as the input to the SEAt module.

[0085] The SEAtt module aggregates the second feature image at different granularities to form image features and extracts them, thus completing the fusion of shallow feature information.

[0086] Specifically, channel attention mechanisms enable neural networks to more effectively extract positional information contained in lower convolutional layers and semantic information from deeper layers. Considering the problem that important feature data is lost as the network depth increases during object detection model training, this invention proposes the idea of ​​fusing shallow and deep feature information based on channel attention mechanisms. This involves giving greater weight to important feature information and discarding unimportant feature information, thereby enhancing the extraction of important features and the understanding of target image feature information. Simultaneously, recognizing the insurmountable information gap between low-dimensional and high-dimensional feature maps, this invention proposes the idea of ​​multi-dimensional feature fusion based on shallow and deep layers. This involves reconstructing semantic and localization information with attention from different dimensions and levels to improve prediction accuracy on feature maps of different sizes.

[0087] Considering the problem that convolutional neural networks lose important data as the number of layers increases, this invention designs a shallow channel attention mechanism (SEAtt). The idea behind SEAtt is to integrate the SENet attention mechanism into the shallow layers of the network. The SENet attention mechanism is a lightweight module that can be applied to any existing network architecture and for any mapping. Taking convolution as an example, the calculation of the convolution kernel directly superimposes all channel information, and all feature maps have the same weight for the classification result. However, in object detection tasks, different target feature extraction networks focus on different key feature regions. If each feature map is given equal importance at the beginning of training, it will increase the time required for network convergence.

[0088] Therefore, the SENet channel attention mechanism is a better choice.

[0089] In the Backbone module of the MCA-CHMR method, the main function is to aggregate shallow images at different granularities and form image features, extracting information from the images for use by subsequent networks. This invention transforms the feature image X (first feature image) output from TCM-MIEM into a feature image E (second feature image) through modules such as CBS and C3_1. The transformed image features are then used as input to the SEAt module.

[0090] Furthermore, when aggregating and extracting image features at different granularities, the process includes:

[0091] The second feature image is transformed by a series of convolution operations to obtain the image features. ;

[0092] Image features are processed using the Squeeze operation in the SENet module. Compression yields a residual channel statistic;

[0093] Based on the residual channel statistics and the prediction of the importance of each channel by the Excitation operation in the SENet module, the nonlinear relationship between different channels is obtained.

[0094] Based on image features The nonlinear relationship between different channels is output through the Scale operation in the SENet module;

[0095] The Excitation operation employs a Sigmoid-based gating mechanism.

[0096] Specifically, the SEAt module assigns greater weight to important features in the feature map and discards unimportant features. It performs three operations on the convolutional feature map: Sequence, Excitation, and Scale. Its structure is as follows: Figure 2 The SEAtt module in [the document] is shown. The specific operation process is as follows:

[0097] First, the input E is transformed through a series of convolution operations to obtain the feature U. Using V = [v1, v2, ..., v...] C Let ] represent a series of convolution kernels, where v C This represents the parameters of the c-th convolutional layer. That is, the output features U = [u1, u2, ..., u...]. C This can be expressed using the following formula 1:

[0098] (1);

[0099] in, This represents the convolution operation. E represents the c-th convolutional kernel of the s-th input. S This represents the s-th input.

[0100] Secondly, the Squeeze operation in the SENet module performs global average pooling on the image features U to compress them into a one-dimensional real number z, representing the residual channel statistics. Assuming the length of the output is set to c, Z... c = [z1, z2......,z c ], (x, y) represents a size of W The characteristic of H is that x is the horizontal coordinate and y is the vertical coordinate. That is, the c-th element of z can be represented by the following formula 2:

[0101] (2);

[0102] Next, the importance of each channel is predicted by the excitation operation in the SENet module, resulting in the nonlinear relationship between different channels. Therefore, a sigmoid-type gating mechanism is used in the excitation operation. Assume... , For two distinct fully connected layers, r represents the dimensionality reduction rate. A smaller r preserves the global information of the previous layer better, but increases computational cost. To achieve a balance between propagation speed and detection accuracy, r in SENet is set to 16. The final output parameter of the Excitation operation is the weight ω for each feature channel, which can be expressed as Equation 3:

[0103] (3);

[0104] Where σ is the Sigmoid function and δ is the ReLU activation function.

[0105] Finally, the obtained 1×1×C real number is combined with the original feature map and output through the Scale operation in the SENet module, as shown in Equation 4:

[0106] (4);

[0107] In the formula Representing image features u c ∈R W×H and scalar ω c The product of corresponding pixels within each channel. The scaling operation allows the network model to automatically learn the importance of each channel, thereby enhancing its ability to recognize image features.

[0108] In summary, the features extracted by the SEAtt method contain more and more accurate auxiliary information, which enhances the model's understanding of the target image's feature information and improves the model's recognition rate.

[0109] In step S2 above, when fusing deep feature information after the Neck layer and before the Prediction layer, the following steps are included:

[0110] By introducing the multidimensional deep channel attention mechanism MCATt into the deep layer of the network, the ECA attention mechanism is reconstructed for semantic and localization information of different dimensions and levels, providing a foundation for the accurate fusion of low, medium and high dimension feature maps at the deep layer of the network.

[0111] By combining the shallow channel attention mechanism SEAtt and the multidimensional deep channel attention mechanism MCAT, a microscopic feature extraction model for multidimensional channel attention is obtained.

[0112] The deep layers of the network include low, medium, and high dimensions.

[0113] Specifically, considering the insurmountable information gap between low-dimensional and high-dimensional feature maps, this invention designs a multi-dimensional deep channel attention mechanism (MCAtt). The idea of ​​MCA is to introduce an ECA attention mechanism into the deep layers of the network (containing low, medium, and high dimensions) to reconstruct the semantic and localization information of different dimensions and levels, providing a foundation for the accurate fusion of low, medium, and high-dimensional feature maps in the deep layers of the network.

[0114] Furthermore, when providing a foundation for the accurate fusion of low, medium, and high-dimensional feature maps at deeper layers of the network, this includes:

[0115] Input low-dimensional features Mid-dimensional characteristics and high-dimensional features After a series of convolution operations, they are transformed into features. ,feature and features ;

[0116] Perform a Squeeze operation to compress features along the spatial dimensions using global average pooling. ,feature and features After compression, one-dimensional real numbers are obtained respectively. One-dimensional real numbers and one-dimensional real numbers ;

[0117] According to real numbers real numbers and real numbers The ECA module uses a one-dimensional convolution with a kernel size of k to perform cross-channel information interaction, and obtains low-dimensional weights, medium-dimensional weights and high-dimensional weights respectively.

[0118] The low-dimensional weights, medium-dimensional weights, and high-dimensional weights are combined with the original feature maps respectively, and the output is obtained through the Scale operation in the ECA module. This provides a foundation for the accurate fusion of feature maps of different dimensions (low, medium, and high).

[0119] Among them, the Neck layer and the Prediction layer each have low-dimensional features. Mid-dimensional characteristics and high-dimensional features The feature output has three different dimensions.

[0120] Specifically, in the Neck module of the MCA-CHMR method, PAN and FPN structures are mainly used to map and superimpose rich shallow location information onto deep features for information fusion. The location information from the lower layers is then passed to higher layers, strengthening the information transfer between different feature maps, accurately preserving spatial information, and thus improving the network's ability to detect large and medium-sized targets. For example... Figure 2 As shown, there are low-level Neck layers leading to Prediction layers. ,middle and high The feature output has three different dimensions. For example, This can be expressed as Formula 5:

[0121] (5);

[0122] Here, + indicates serial processing of the feature image, ⊕ indicates tensor concatenation, and its function is to assign different weights to input features at different levels. f represents the processing of the feature image by the SPPF module, g represents the processing of the feature image by the CBS module, and d represents the processing of the feature image by the C3_1_F module;

[0123] The ECA attention mechanism performs three operations on the feature map obtained from convolution: Sequence, Conv, and Scale. Its structure is as follows: Figure 2 The MCATT module is shown below. The specific processing procedure of the MCATT module is as follows:

[0124] First, enter E h Feature U is obtained through a series of convolution operations and transformations. h .

[0125] Secondly, a Squeeze operation is performed, utilizing global average pooling to compress the feature U along the spatial dimension.h After compression, it becomes a one-dimensional real number z. The calculation for the c-th cell in z is shown in Formula 6:

[0126] (6);

[0127] Next, to avoid dimensionality reduction, the ECA module uses a one-dimensional convolution with a kernel size of k to achieve cross-channel information interaction, as shown in Equation 7:

[0128] (7);

[0129] Wherein, C1D is a one-dimensional convolution, and k is the size of the one-dimensional convolution kernel, used to represent the range of cross-channel interaction. k has a feature mapping relationship with the number of channels c, which can be adaptively calculated using the following formula 8.

[0130] (8);

[0131] Where, |n| odd Let γ be the closest odd number to n, and let b be 2 and 1 respectively. By mapping ψ, higher-dimensional channels have longer interactions, while lower-dimensional channels have shorter interactions through the use of nonlinear mapping.

[0132] Finally, the output weights are combined with the original feature map and the output is processed through the Scale operation in the ECA module. The final residual feature representation is shown in Equation 9:

[0133] (9);

[0134] Similarly, low-dimensional residual characteristics can be obtained. and the characteristics of medium-dimensional residuals .

[0135] In summary, the features extracted by the MCATt method contain effective auxiliary information of different dimensions and levels, providing a foundation for the accurate fusion of low, medium, and high-dimensional feature maps, enhancing the model's understanding of target image feature information, and improving the model's prediction accuracy on feature maps of different dimensions and sizes.

[0136] In step S4 above, when generating prediction boxes through the Prediction layer and using these prediction boxes to predict the features of the microscopic image of traditional Chinese medicine, such as... Figure 2 As shown, it includes:

[0137] By using a microscopic identification model for traditional Chinese medicine, features from different receptive fields are fused in three different dimensions (low, medium, and high) to generate feature maps at three different target scales.

[0138] We use feature maps at three different target scales to predict targets in microscopic images of traditional Chinese medicine.

[0139] Furthermore, when predicting targets in microscopic images of traditional Chinese medicine, the following are included:

[0140] The input microscopic image of Chinese medicine is divided into grids to obtain the grids where the target is located, and the target is predicted using the grids.

[0141] During the training phase, a three-part loss method consisting of target loss, category loss, and confidence loss is used to predict microscopic images of traditional Chinese medicine.

[0142] Non-maximum suppression is used to remove redundant prediction boxes, select high-quality detection results, and obtain the characteristics of cells in the microscopic images of traditional Chinese medicine.

[0143] Specifically, the prediction process is as follows:

[0144] First, the input microscopic image of traditional Chinese medicine is divided into a grid. If the center of a target exists within the grid, the target is predicted using that grid. Each grid cell's prediction includes the location information of three target bounding boxes and one confidence score. A target bounding box corresponds to four location information points (x, y, w, h) and one confidence score. Here, x and y represent the center point position of the object; w and h represent the distance from the center point to the two sides of the object, respectively; the confidence score C represents both the confidence that the predicted target bounding box contains the target and the accuracy of the target bounding box prediction, as shown in Formula 10.

[0145] (10);

[0146] Where IOU = (A∩B) / (A∪B), A represents the ground truth bounding box, and B represents the predicted bounding box. When P r When (obj)=1, it indicates that a target exists in the image; when P r When (obj)=0, it indicates that there is no target in the image.

[0147] Secondly, a three-part loss system—target loss, category loss, and confidence loss—is used to predict the microscopic images of traditional Chinese medicine during the training phase. The target loss is calculated using binary cross-entropy, as shown in Equation 11:

[0148] (11);

[0149] Where S×S represents dividing the input image into an S×S mesh; N represents the number of prediction boxes that one mesh is responsible for; This indicates the position information of the actual bounding box; This indicates the location information of the prediction box; The value is 1 if the j-th prediction box of the i-th network is responsible for predicting object obj, and 0 otherwise; the classification loss is also calculated using binary cross-entropy, as shown in Formula 12:

[0150] (12);

[0151] Where c represents the number of categories; p i (c) represents the probability of the true class; Confidence loss uses the probability of the predicted class.

[0152] The confidence loss CIOU is calculated as shown in Formula 13:

[0153] (13);

[0154] in, λ represents 0 if the j-th prediction box of the i-th network is not responsible for predicting objects, and 1 otherwise; noobj To reduce the confidence loss of the predicted bounding box for a non-existent object obj, λ is... noobj Set to 0.5;

[0155] The total loss is the weighted sum of the target loss, category loss, and confidence loss, as shown in Formula 14:

[0156] (14);

[0157] Finally, non-max suppression (NMS) is used to remove redundant prediction boxes and select high-quality detection results.

[0158] The above process is used to obtain the characteristics of cells in microscopic images of traditional Chinese medicine, thereby enabling the identification of cells in microscopic images of traditional Chinese medicine.

[0159] The present invention provides a traditional Chinese medicine (TCM) microscopic recognition system based on a multi-channel attention mechanism, comprising an enhancement unit, an auxiliary information acquisition unit, and a prediction unit. The enhancement unit, at the Input layer, combines the characteristics of incomplete cell structure, three-dimensional image, and uneven feature distribution in TCM microscopic images to obtain a TCM microscopic image data enhancement model, and uses this model to enhance the TCM microscopic image data. The auxiliary information acquisition unit fuses shallow feature information at the Backbone layer and deep feature information after the Neck layer and before the Prediction layer to obtain a multi-channel attention-based microscopic feature extraction model; this model extracts implicit and effective auxiliary information from different channels from shallow to deep. The prediction unit utilizes this implicit and effective auxiliary information to generate prediction boxes through the Prediction layer, and uses these prediction boxes to predict the features of the TCM microscopic image, thereby obtaining the cell features and completing the cell recognition of the TCM microscopic image.

[0160] Example

[0161] 1. Experimental Setup

[0162] 1.1 Dataset

[0163] As shown in Table 1, two powdered traditional Chinese medicines, Scutellaria baicalensis and Magnolia officinalis, were prepared into glass slides, and microscopic images of the traditional Chinese medicines were acquired using a Nikon E200 electron microscope with a 40 / 0.65 objective lens. The two traditional Chinese medicines involved nine cell features, resulting in a small sample dataset of 11,060 microscopic images and 12,840 labeled data. Furthermore, in this embodiment, labelImg was used for cell labeling of the traditional Chinese medicine microscopic images, and the ratio of the training dataset, validation dataset, and test dataset was 8:1:1.

[0164] Table 1. Microscopic Chinese Medicine Image Annotation Dataset

[0165]

[0166] 1.2 Evaluation Indicators

[0167] To comprehensively evaluate the MCA-CHMR method proposed in this invention, precision, recall, AP curve, and mean average precision (MAP) were selected as evaluation metrics.

[0168] Precision represents the ratio of true positive instances predicted as true to all predicted positive instances. It is calculated as shown in Formula 15:

[0169] (15);

[0170] Where TP indicates that the predicted value is the same as the true value, and the predicted value is a positive sample; FP indicates that the predicted value is different from the true value, and the predicted value is a positive sample. Recall represents the ratio of the true positive examples predicted as true to all true positive examples. The calculation method is shown in Formula 16:

[0171] (16);

[0172] In this context, TP indicates that the predicted value is the same as the true value, and the predicted value is a positive sample; FN indicates that the predicted value is different from the true value, and the predicted value is a negative sample.

[0173] The AP curve is the area enclosed by the curves representing Precision and Recall. Generally, a lower Recall corresponds to a higher Precision, and vice versa. In other words, a larger AP curve indicates better model performance.

[0174] MAP is a comprehensive evaluation metric that focuses on sequence weights and has become one of the most important practical metrics for object detection in recent years. mAP@.0.5 means that when IoU=0.5, the AP of all images in each class is calculated and then averaged over all classes. The higher the value of mAP@.0.5, the better.

[0175] 1.3 Parameter Optimization

[0176] To determine the optimal solution during model training, four important parameters were considered to obtain the optimal MAP value within a learning timeframe. These four parameters are: learning rate (lrf), weight decay, classification loss coefficient (cls), and confidence loss coefficient (obj). The learning rate (lrf) is a hyperparameter that guides how the network weights are adjusted using the gradient of the loss function. It determines the scale of network parameter updates; if the slope is set too small, more iterations are needed until convergence. Its value is typically set for 5-10 training iterations, exploring a range of 1-2, with each exploration increment being a factor of 10 (expanding or shrinking). Weight decay, also known as L2 regularization, aims to decay the weights to smaller values, reducing overfitting to some extent. Its value starts at a very small value, explores with a scaling factor of 5-10, and is adjusted to a very small value (e.g., 0.01) after obtaining a suitable magnitude. These four parameters can vary within the range of 0 and 1.

[0177] In this embodiment, the four hyperparameters affecting network performance are set to four levels. If a grid search method is used, a total of 256(4) possible combinations need to be analyzed. It is conceivable that so many experiments will consume a lot of time and require more computing resources. In order to reduce the number of experiments and obtain a relatively good experimental result, this invention introduces the Taguchi method, which is an industrial optimization method that finds the optimal structure through orthogonal combinations of influencing factors. Due to the existence of orthogonal arrays, this method reduces the number of experiments, and therefore, it is suitable for studies that cannot cover all possible experiments.

[0178] The four parameters and their levels proposed in this embodiment are shown in Table 2.

[0179] Table 2 Experimental parameters and their levels

[0180]

[0181] Table 3 presents a 20-row orthogonal array, where the four parameters are interconnected and interact according to the levels shown in Table 2. The last three columns of Table 3 represent the mAP@.0.5 values ​​for *Scutellaria baicalensis* Fibres, *Scutellaria baicalensis* Stonecells, and *Scutellaria baicalensis* Oilcells obtained in each experiment. Table 3 shows that the orthogonal array in row 17 corresponds to the highest mAP@.0.5 values ​​for *Scutellaria baicalensis* Fibres, *Scutellaria baicalensis* Stonecells, and *Scutellaria baicalensis* Oilcells. Therefore, the optimal mAP@.0.5 value is achieved when the learning rate (lrf), weight decay (L2), classification loss coefficient (cls), and confidence loss coefficient (obj) are 0.01, 0.00036, 0.5, and 1.0, respectively.

[0182] Table 3 Orthogonal Experiment Table

[0183]

[0184] 1.4 Comparison Methods

[0185] To fully verify the microscopic identification method of traditional Chinese medicine of the present invention, this embodiment selected two currently popular one-stage object detection algorithms (YOLOv5 and SSD) and two traditional two-stage object detection algorithms (Faster RCNN and RetinaNet) for comparative experiments.

[0186] Single Shot MultiBox Detector (SSD) is an algorithm based on a feedforward convolutional neural network. It classifies and locates objects using regression and combines the anchor point concept from Faster RCNN for prediction. The prior boxes in the algorithm enable the model to converge faster and reduce training costs.

[0187] Faster R-CNN is a true end-to-end deep learning detection algorithm. Its biggest innovation lies in adding an RPN network to generate candidate boxes based on the Anchor mechanism. Ultimately, it integrates feature extraction, candidate box selection, bounding box regression, and classification into a single network, thereby effectively improving detection accuracy and efficiency.

[0188] The core of the RetinaNet network is Focal Loss, which surpasses the accuracy of Two-Stage networks and the speed of One-Stage networks, achieving a comprehensive improvement over Two-Stage networks for the first time.

[0189] Comparative experiments were conducted to identify cells in microscopic images of nine Chinese medicinal herbs, including Scutellaria baicalensis and Magnolia officinalis, and good results were achieved. The comparative experiments used the same dataset for training, with identical hardware and software environments, model training parameters, and number of iterations. The mAP@.0.5 results for the comparative experiments are shown in Table 4.

[0190] Table 4 Comparison of Training Results for Different Convolutional Neural Networks

[0191]

[0192] The comparison results show that, for the classification of cells in the microscopic images of the three kinds of Chinese herbal medicines, including Magnolia officinalis, the Chinese herbal medicine microscopic image recognition method proposed in this paper has the best performance on mAP@.0.5, which shows that the method is feasible and more robust in solving the problem of classifying cells in microscopic images of Chinese herbal medicines.

[0193] 2. Data Augmentation Validation

[0194] Compared to RetinaNet, RCNN, and SSD, YOLOv5 has a more streamlined network structure, the highest feature extraction accuracy, and a faster detection speed. Therefore, in this embodiment, the detection results based on YOLOv5 are used as a comparative experiment to verify the accuracy of the proposed Traditional Chinese Medicine Microscopic Image Data Augmentation Model (TCM-MIEM). The YOLOv5 pre-training image library uses a dataset of 9793 TCM microscopic images; the data augmentation images are mainly 5000 images randomly selected from the dataset. Figure 4The results show the experimental results for Scutellaria baicalensis Corkcells (550 images) and Scutellaria baicalensis Vessels (299 images). The results indicate that after adding the data augmentation model, the accuracy and recall of Scutellaria baicalensis Vessels are improved, and the effect is more significant than that of Scutellaria baicalensis Corkcells. This demonstrates that the data augmentation method proposed in this invention is effective even for problems with extremely small sample sizes. Therefore, the TCM-MIEM proposed in this invention is very helpful in improving the performance of cell recognition in microscopic images of traditional Chinese medicine.

[0195] 3. Validation of attention effectiveness

[0196] To verify the effectiveness of the multi-channel attention microscopic feature extraction model, this paper conducts comparative experiments on the YOLOv5 network model, SEAtt, MCATt, and the multi-channel attention microscopic feature extraction model (SEAtt+MCAtt), analyzing the target recognition performance of Scutellaria baicalensis Corkcells and Magnolia officinalis Stonecells. The experimental results are as follows: Figure 5 As shown.

[0197] Experimental results show that, in terms of accuracy, incorporating the attention mechanism into Scutellaria baicalensis Corkcells yields the most significant improvement, with SEAt increasing accuracy by 1.8%, MCAT increasing it by 5.5%, and the combined SEAt+MCCAT improving it by 8.1%. Magnolia officinalis Stonecells also showed improvements of 5%, 3.3%, and 0.4%, respectively. Therefore, this indicates that the multi-channel attention-based microscopic feature extraction model is highly helpful in improving the performance of cell recognition in microscopic images of traditional Chinese medicine.

[0198] 4. Performance Comparison Experiment of Various Microscopic Identification Methods for Traditional Chinese Medicine

[0199] To verify the effectiveness of the traditional Chinese medicine (TCM) microscopic image data enhancement model and the multi-channel attention microscopic feature extraction model, this paper conducted comparative experiments on the YOLOv5 network model, the TCM microscopic image data enhancement model (TCM-MIEM), the multi-channel attention microscopic feature extraction model (SEAtt+MCAtt), and the fused TCM microscopic recognition model (MCA-CHMR). The target recognition performance of Scutellaria baicalensis Corkcells and Magnolia officinalis Stonecells cells was analyzed. The experimental results are as follows: Figure 6 As shown. By Figure 6As can be seen, compared with the YOLOv5 network model, the fused microscopic recognition model for traditional Chinese medicine showed an 8.1% improvement in prediction precision and a 1.7% improvement in mAP@.5 for Scutellaria baicalensis (Corkcells). For Magnolia officinalis (Stonecells), the prediction precision improved by 8.4% and the mAP@.5 improved by 0.7%. Furthermore, the AP curves formed by the precision and recall results of YOLOv5, TCM-MIEM, SEAt+MCAtt, and MCA-CHMR are shown below. Figure 7 As shown.

[0200] Depend on Figure 7 It can be seen that the areas of precision and recall of Scutellaria baicalensis Corkcells in YOLOv5, TCM-MIEM, SEAtt+MCAtt, and MCA-CHMR experiments were 70.488%, 74.11%, 71.72%, and 77.41%, respectively, indicating that the MCA-CHMR proposed in this invention is of great help in improving the performance of cell recognition in microscopic images of traditional Chinese medicine.

[0201] 5. Data Quality Assessment and Analysis

[0202] In the comparative experiments of YOLOv5, TCM-MIEM, SEAtt+MCAtt, and MCA-CHMR, it can be seen that the mAP@.5 value of Houpu Fibres, which has the largest amount of data in the small sample dataset, hardly changed in each experiment. Figure 8 As shown in Table 5, the AP curve remained almost unchanged. This phenomenon contradicts the results of other microscopic images of Chinese herbal medicine cells in this experiment. Further analysis revealed that 4540 out of the 7555 images in the Magnolia officinalis Fibre dataset were not clear enough due to unfamiliarity with microscope operation and poor control of reagent ratios during slide preparation. Therefore, this demonstrates that data quality is another effective way to improve the recognition rate.

[0203] Table 5. Comparative experiments on AP values ​​of Magnolia officinalis Fibre

[0204]

[0205] This invention addresses the problem that existing image recognition methods cannot be directly applied to TCM microscopic images due to their small size and uneven distribution. It proposes a TCM microscopic recognition method based on a multi-channel attention mechanism. First, considering the incomplete cell structure, three-dimensionality, and uneven feature distribution of TCM microscopic images, a TCM microscopic image data enhancement model is designed to expand and balance existing TCM microscopic image datasets. Second, by fusing shallow and deep (including low, medium, and high dimensions) feature information, a multi-channel attention microscopic feature extraction model is proposed, extracting more representative auxiliary information from small-scale datasets. Experimental results demonstrate that the proposed TCM microscopic recognition method based on a multi-channel attention mechanism can effectively extract salient targets from TCM microscopic images, proving the effectiveness and robustness of the method provided.

[0206] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-substantial changes and substitutions made by those skilled in the art based on the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. A method for microscopic identification of traditional Chinese medicine based on a multidimensional channel attention mechanism, characterized in that, Includes the following steps: By combining the characteristics of incomplete cell structure, three-dimensional image, and uneven feature distribution of traditional Chinese medicine microscopic images in the Input layer, a traditional Chinese medicine microscopic image data enhancement model is obtained, and the traditional Chinese medicine microscopic image data enhancement model is used to enhance the traditional Chinese medicine microscopic image data. Shallow feature information is fused in the Backbone layer, and deep feature information is fused after the Neck layer and before the Prediction layer to obtain a microscopic feature extraction model with multi-channel attention. The microscopic feature extraction model of the multidimensional channel attention is used to obtain implicit and effective auxiliary information from different channels from shallow to deep layers. Using the implicit and effective auxiliary information, a prediction box is generated through the Prediction layer, and the prediction box is used to predict the features of the microscopic image of traditional Chinese medicine to obtain the features of the cells in the microscopic image of traditional Chinese medicine, thus completing the identification of the cells in the microscopic image of traditional Chinese medicine. Specifically, when fusing shallow feature information in the Backbone layer, the process includes: incorporating the SENet attention mechanism into the shallow layer of the network to obtain the shallow channel attention mechanism SEAt; transforming the first feature image output after enhancement by the traditional Chinese medicine microscopic image data enhancement model into a second feature image after processing by the feature processing module, and using the second feature image as the input to the SEAt module; and using the SEAt module to aggregate the second feature image at different granularities to form image features and extract them, thereby completing the fusion of shallow feature information. When fusing deep feature information after the Neck layer and before the Prediction layer, the following are included: By introducing the multidimensional deep channel attention mechanism MCATt into the deep layer of the network, the ECA attention mechanism is reconstructed for semantic and localization information of different dimensions and levels, providing a foundation for the accurate fusion of low, medium and high dimension feature maps at the deep layer of the network. By combining the shallow channel attention mechanism SEAtt and the multidimensional deep channel attention mechanism MCAT, a microscopic feature extraction model for the multidimensional channel attention is obtained. The deep layers of the network include low, medium, and high dimensions.

2. The method for microscopic identification of traditional Chinese medicine based on multidimensional channel attention mechanism according to claim 1, characterized in that, When enhancing the data using the aforementioned traditional Chinese medicine microscopic image data enhancement model, the following steps are included: Randomly select n images from the original training dataset, and perform horizontal average segmentation on the n images to obtain several segmented images; From the aforementioned segmented images, n images are randomly selected and horizontally stitched together to obtain several stitched images. A new training dataset is constructed using the stitched images and the images in the original training dataset.

3. The method for microscopic identification of traditional Chinese medicine based on multidimensional channel attention mechanism according to claim 2, characterized in that, When constructing a completely new training dataset, the following are included: The stitched images and the images in the original dataset are then subjected to mirroring, translation, and rotation processing.

4. The method for microscopic identification of traditional Chinese medicine based on a multidimensional channel attention mechanism according to claim 1, characterized in that, When aggregating and extracting image features at different granularities, the following steps are included: The second feature image is transformed by a series of convolution operations to obtain the image features. ; The image features are processed using the Squeeze operation in the SENet module. Compression yields a residual channel statistic; Based on the residual channel statistics, and by predicting the importance of each channel using the Excitation operation in the SENet module, the nonlinear relationship between different channels is obtained. Based on the image features The nonlinear relationship between the different channels is output through the Scale operation in the SENet module; The Excitation operation employs a Sigmoid-based gating mechanism.

5. The method for microscopic identification of traditional Chinese medicine based on a multidimensional channel attention mechanism according to claim 1, characterized in that, When providing a foundation for the accurate fusion of low, medium, and high-dimensional feature maps at deeper layers of the network, this includes: Input low-dimensional features Mid-dimensional characteristics and high-dimensional features After a series of convolution operations, they are transformed into features. ,feature and features ; Perform a Squeeze operation to compress the features along the spatial dimension using global average pooling. The features and the aforementioned features After compression, one-dimensional real numbers are obtained respectively. One-dimensional real numbers and one-dimensional real numbers ; According to the real number The real number and the real number The ECA module uses a one-dimensional convolution with a kernel size of k to perform cross-channel information interaction, and obtains low-dimensional weights, medium-dimensional weights and high-dimensional weights respectively. The low-dimensional weights, medium-dimensional weights, and high-dimensional weights are combined with the original feature maps and output through the Scale operation in the ECA module to obtain low-dimensional residual features, medium-dimensional residual features, and high-dimensional residual features, respectively, providing a foundation for the accurate fusion of low, medium, and high-dimensional feature maps. The Neck layer and the Prediction layer each have low-dimensional features. Mid-dimensional characteristics and high-dimensional features The feature output has three different dimensions.

6. The method for microscopic identification of traditional Chinese medicine based on multidimensional channel attention mechanism according to claim 1, characterized in that, When generating prediction boxes through the Prediction layer and using these prediction boxes to predict the features of microscopic images of traditional Chinese medicine, the process includes: By using a microscopic identification model for traditional Chinese medicine, features from different receptive fields are fused in three different dimensions (low, medium, and high) to generate feature maps at three different target scales. The feature maps of the three different target scales are used to predict targets in microscopic images of traditional Chinese medicine.

7. The method for microscopic identification of traditional Chinese medicine based on a multi-dimensional channel attention mechanism according to claim 1, characterized in that, Predicting targets in microscopic images of traditional Chinese medicine includes: The input microscopic image of Chinese medicine is divided into grids to obtain the grids where the center of the target exists, and the target is predicted using the grids. The microscopic images of traditional Chinese medicine are predicted during the training phase using a loss consisting of three parts: target loss, category loss, and confidence loss. Non-maximum suppression is used to remove redundant prediction boxes, select high-quality detection results, and obtain the characteristics of cells in the microscopic images of traditional Chinese medicine.

8. A microscopic recognition system for traditional Chinese medicine based on a multidimensional channel attention mechanism, characterized in that, include: Enhancement unit: used to combine the characteristics of incomplete cell structure, three-dimensional image, and uneven feature distribution of traditional Chinese medicine microscopic images in the Input layer to obtain a traditional Chinese medicine microscopic image data enhancement model, and to enhance the traditional Chinese medicine microscopic image data using the traditional Chinese medicine microscopic image data enhancement model; Auxiliary information acquisition unit: used to fuse shallow feature information in the Backbone layer and deep feature information after the Neck layer and before the Prediction layer to obtain a multi-dimensional channel attention microscopic feature extraction model; through the multi-dimensional channel attention microscopic feature extraction model, implicit effective auxiliary information is obtained from different channels from shallow to deep layers; Prediction unit: Used to generate a prediction box through the Prediction layer using the implicit effective auxiliary information, and use the prediction box to predict the features of the microscopic image of Chinese medicine, thereby obtaining the features of the cells in the microscopic image of Chinese medicine and completing the identification of the cells in the microscopic image of Chinese medicine. Specifically, when fusing shallow feature information in the Backbone layer, the process includes: incorporating the SENet attention mechanism into the shallow layer of the network to obtain the shallow channel attention mechanism SEAt; transforming the first feature image output after enhancement by the traditional Chinese medicine microscopic image data enhancement model into a second feature image after processing by the feature processing module, and using the second feature image as the input to the SEAt module; and using the SEAt module to aggregate the second feature image at different granularities to form image features and extract them, thereby completing the fusion of shallow feature information. When fusing deep feature information after the Neck layer and before the Prediction layer, the following are included: By introducing the multidimensional deep channel attention mechanism MCATt into the deep layer of the network, the ECA attention mechanism is reconstructed for semantic and localization information of different dimensions and levels, providing a foundation for the accurate fusion of low, medium and high dimension feature maps at the deep layer of the network. By combining the shallow channel attention mechanism SEAtt and the multidimensional deep channel attention mechanism MCAT, a microscopic feature extraction model for the multidimensional channel attention is obtained. The deep layers of the network include low, medium, and high dimensions.

Images