CN-116524495-B - Traditional Chinese medicine microscopic identification method and system based on multidimensional channel attention mechanism

CN116524495BCN 116524495 BCN116524495 BCN 116524495BCN-116524495-B

Abstract

The invention discloses a traditional Chinese medicine microscopic identification method and system based on a multidimensional channel attention mechanism, comprising the steps of combining the characteristics of incomplete cell structure, stereoscopic image and unbalanced characteristic distribution of a traditional Chinese medicine microscopic image in an Input layer to obtain a traditional Chinese medicine microscopic image data enhancement model, and enhancing traditional Chinese medicine microscopic image data by utilizing the enhancement model; the method comprises the steps of merging shallow characteristic information in a backstone layer, merging deep characteristic information in front of a Neck-layer rear Prediction layer to obtain a multi-dimensional channel attention microscopic characteristic extraction model, obtaining hidden effective auxiliary information from different channels from the shallow layer to the deep layer through the extraction model, predicting Chinese medicine microscopic image characteristics by utilizing the hidden effective auxiliary information through a Prediction layer generation Prediction frame to obtain Chinese medicine microscopic image cell characteristics, and completing recognition of Chinese medicine microscopic image cells, so that the Prediction accuracy of a Chinese medicine microscopic image is improved.

Inventors

ZHU XIAOYING
PANG GUANGYAO
YU ZHENMING
GONG PING
Lu keda
DENG JIAWEI
Zhong Wenrui
ZENG QINGHU
CHEN YIFENG

Assignees

梧州学院

Dates

Publication Date: 20260512
Application Date: 20230214

Claims (8)

1. A traditional Chinese medicine microscopic identification method based on a multidimensional channel attention mechanism is characterized by comprising the following steps: Combining the characteristics of incomplete cell structure, stereo image and unbalanced characteristic distribution of the traditional Chinese medicine microscopic image in the Input layer to obtain a traditional Chinese medicine microscopic image data enhancement model, and enhancing the traditional Chinese medicine microscopic image data by utilizing the traditional Chinese medicine microscopic image data enhancement model; Fusing shallow characteristic information at a back bone layer, and fusing deep characteristic information before a pre-section layer behind Neck layers to obtain a microscopic characteristic extraction model of the attention of a multidimensional channel; obtaining hidden effective auxiliary information from different channels from a shallow layer to a deep layer through a microscopic feature extraction model of the multi-dimensional channel attention; Generating a Prediction frame through a Prediction layer by utilizing the implicit effective auxiliary information, and predicting the characteristics of the traditional Chinese medicine microscopic image by utilizing the Prediction frame to obtain the characteristics of the traditional Chinese medicine microscopic image cells so as to finish the identification of the traditional Chinese medicine microscopic image cells; the method comprises the steps of merging shallow characteristic information in a backstone layer, merging SENet attention mechanisms in a shallow layer of a network to obtain a shallow channel attention mechanism SEAtt, processing a first characteristic image which is enhanced by the traditional Chinese medicine microscopic image data enhancement model and then output, converting the first characteristic image into a second characteristic image by a characteristic processing module, and taking the second characteristic image as input of a SEAtt module, and carrying out aggregation of different granularities on the second characteristic image by the SEAtt module to form image characteristics and extracting to finish the merging of the shallow characteristic information; When the feature information of the deep layer is fused before the pre-section layer after Neck layers, the method comprises the following steps: Introducing an ECA attention mechanism into the deep layer of the network through a multidimensional depth channel attention mechanism MCAtt, reconstructing the attention of semantic information and positioning information of different dimensions and different layers, and providing a basis for the accurate fusion of feature graphs of different dimensions of low, medium and high in the deep layer of the network; Combining the shallow channel attention mechanism SEAtt and the multidimensional depth channel attention mechanism MCAt to obtain a microscopic feature extraction model of the multidimensional channel attention; Wherein the deep layer of the network comprises low, medium and high different dimensions.
2. The multi-dimensional channel attention mechanism based traditional Chinese medicine microscopic identification method according to claim 1, wherein when the traditional Chinese medicine microscopic image data enhancement model is utilized to enhance data, the method comprises the following steps: Randomly selecting n pictures from an original training data set, and carrying out average segmentation on the n pictures in the horizontal direction to obtain a plurality of segmented pictures; Randomly extracting n pictures from the plurality of segmented pictures to splice in the horizontal direction, and obtaining a plurality of spliced pictures; and constructing a brand new training data set by utilizing the spliced pictures and the pictures in the original training data set.
3. The method for microscopic recognition of traditional Chinese medicine based on multidimensional channel attention mechanism according to claim 2, wherein when constructing a completely new training data set, the method comprises: And carrying out mirror image processing, translation processing and rotation processing on the spliced pictures and the pictures in the original data set.
4. The method for microscopic recognition of traditional Chinese medicine based on multidimensional channel attention mechanism according to claim 1, wherein when aggregation of different granularities is performed to form image features and extraction is performed, the method comprises: transforming the second characteristic image after a series of convolution operations to obtain image characteristics ; Characterizing the image by a Squeeze operation in SENet modules Compressing to obtain a residual channel statistic; Predicting the importance of each channel by using the accounting operation in the SENet module according to the residual channel statistics to obtain nonlinear relations of different channels; According to the image characteristics And the nonlinear relation of the different channels is output through Scale operation in the SENet module; wherein, the gating mechanism in the form of Sigmoid is adopted in the specifying operation.
5. The multi-dimensional channel attention mechanism-based traditional Chinese medicine microscopic identification method according to claim 1, wherein when the deep layer of the network provides a basis for accurate fusion of low, medium and high different dimension feature maps, the method comprises the following steps: low-dimensional features to be input Middle dimension feature High-dimensional features Transformed into features after a series of convolution operations Features and characteristics Features and characteristics ; Performing a Squeeze operation, compressing the features along a spatial dimension using global averaging pooling The features are The features After compression, one-dimensional real numbers are respectively obtained Real number in one dimension Real number in one dimension ; According to the real number Said real number Said real number The ECA module performs cross-channel information interaction through one-dimensional convolution with a convolution kernel of k to respectively obtain low-dimensional weight, medium-dimensional weight and Gao Weiquan weight; Combining the low-dimensional weight, the medium-dimensional weight and the Gao Weiquan weight with original feature images respectively, outputting the low-dimensional weight, the medium-dimensional weight and the Gao Weiquan weight through Scale operation in the ECA module to obtain a low-dimensional residual feature, a medium-dimensional residual feature and a high-dimensional residual feature respectively, and providing a basis for accurate fusion of low, medium and high different-dimensional feature images; Wherein the Neck layers respectively have low-dimensional characteristics towards the Prediction layer Middle dimension feature High-dimensional features Three different dimensions of feature output.
6. The method for microscopic recognition of traditional Chinese medicine based on multidimensional channel attention mechanism according to claim 1, wherein when generating a Prediction frame through a Prediction layer and predicting the characteristics of a traditional Chinese medicine microscopic image by using the Prediction frame, the method comprises the steps of: the characteristics of different receptive fields are fused in low, medium and high 3 different dimensions through a traditional Chinese medicine microscopic identification model respectively, so that three characteristic diagrams with different target dimensions are generated; and predicting the target in the traditional Chinese medicine microscopic image by utilizing the characteristic diagrams of the three different target scales.
7. The method for microscopic recognition of chinese medicine based on multidimensional channel attention mechanism of claim 1, wherein in predicting the target in the chinese medicine microscopic image, comprising: dividing an input traditional Chinese medicine microscopic image into grids, obtaining grids with centers of targets, and predicting the targets by utilizing the grids; Predicting the traditional Chinese medicine microscopic image in a training stage by adopting the loss of three parts of target loss, category loss and confidence loss; and removing redundant prediction frames by adopting non-maximum suppression, screening out high-quality detection results, and obtaining the characteristics of the traditional Chinese medicine microscopic image cells.
8. A multi-dimensional channel attention mechanism-based traditional Chinese medicine microscopic identification system, comprising: The enhancement unit is used for combining the characteristics of incomplete cell structure, image stereo and unbalanced characteristic distribution of the traditional Chinese medicine microscopic image in the Input layer to obtain a traditional Chinese medicine microscopic image data enhancement model, and enhancing the traditional Chinese medicine microscopic image data by utilizing the traditional Chinese medicine microscopic image data enhancement model; The auxiliary information acquisition unit is used for fusing the characteristic information of the shallow layer in the back bone layer and fusing the characteristic information of the deep layer in front of the pre-section layer behind the Neck layers to obtain a multi-dimensional channel attention microscopic characteristic extraction model; the Prediction unit is used for generating a Prediction frame through the Prediction layer by utilizing the implicit effective auxiliary information, predicting the characteristics of the traditional Chinese medicine microscopic image by utilizing the Prediction frame to obtain the characteristics of the traditional Chinese medicine microscopic image cells, and completing the identification of the traditional Chinese medicine microscopic image cells; the method comprises the steps of merging shallow characteristic information in a backstone layer, merging SENet attention mechanisms in a shallow layer of a network to obtain a shallow channel attention mechanism SEAtt, processing a first characteristic image which is enhanced by the traditional Chinese medicine microscopic image data enhancement model and then output, converting the first characteristic image into a second characteristic image by a characteristic processing module, and taking the second characteristic image as input of a SEAtt module, and carrying out aggregation of different granularities on the second characteristic image by the SEAtt module to form image characteristics and extracting to finish the merging of the shallow characteristic information; When the feature information of the deep layer is fused before the pre-section layer after Neck layers, the method comprises the following steps: Introducing an ECA attention mechanism into the deep layer of the network through a multidimensional depth channel attention mechanism MCAtt, reconstructing the attention of semantic information and positioning information of different dimensions and different layers, and providing a basis for the accurate fusion of feature graphs of different dimensions of low, medium and high in the deep layer of the network; Combining the shallow channel attention mechanism SEAtt and the multidimensional depth channel attention mechanism MCAt to obtain a microscopic feature extraction model of the multidimensional channel attention; Wherein the deep layer of the network comprises low, medium and high different dimensions.

Description

Traditional Chinese medicine microscopic identification method and system based on multidimensional channel attention mechanism Technical Field The invention relates to the technical field of microscopic image information processing, in particular to a traditional Chinese medicine microscopic identification method and system based on a multidimensional channel attention mechanism. Background With the vigorous development of the traditional Chinese medicine market, the traditional Chinese medicine detection and identification business also develops well. The current medicinal materials have few wild herbs, more cultivated herbs and cell mutation phenomena, are identified according to experience and microscopic patterns, and have low identification rate and low specificity. In addition, because of the lack of a complete cell data comparison library, the traditional microscopic map and other data are scattered and incomplete, the updating iteration is slow, and subjective factors for describing the target are more. The traditional Chinese medicinal material identification method mainly comprises basic source identification, character identification, microscopic identification and physicochemical identification. The microscopic identification is based on the principle that microscopic characteristics of different medicinal materials are different, and the microscopic observation of the tissue structure and the powder characteristics of the medicinal materials is utilized to realize the true and false identification. Currently, microscopic identification has become a major means of identification for pharmaceutical enterprises, hospitals and drug monitoring institutions as the cost of microscopes has decreased year by year. In recent years, many image recognition methods based on deep learning are formed thanks to the development of artificial neural networks. Image recognition (also known as object detection) methods have been widely used to accurately detect and locate objects in images by learning from data by building neural network models. The target detection algorithm based on deep learning is mainly divided into Two types, namely a Two-Stage detection algorithm based on regional suggestion and an One-Stage detection algorithm based on regression thought. Compared with the Two-Stage detection algorithm, the One-Stage detection algorithm has higher speed, and the represented algorithm is SSD, retinaNet, YOLO series of algorithms. The target detection algorithm based on deep learning is widely applied to various industries, in particular to a YOLO series algorithm in One-Stage detection algorithm. In the YOLO series algorithm, the YOLOv network is easier to put into production due to the use of the Pytorch framework, and has the advantages of being friendly to users, convenient to train, fast in training speed, high in accuracy and the like, and is widely used. However, since YOLOv network has poor detection effect on small targets, the following problems exist when predicting traditional Chinese medicine microscopic images: (1) Because of the problems of incomplete structure, stereo images, unbalanced characteristic distribution and the like of the traditional Chinese medicine microscopic images, the prediction accuracy is lower; (2) When the traditional Chinese medicine microscopic image is identified, important data are lost along with the deepening of the layer number of the convolutional neural network, and an information gap exists between the low-dimensional characteristic image and the high-dimensional characteristic image, so that the prediction accuracy is further reduced. Disclosure of Invention In order to overcome the defects of the prior art, the invention provides a traditional Chinese medicine microscopic identification method and system based on a multidimensional channel attention mechanism, which are used for solving the technical problem that the traditional microscopic image identification technology is low in accuracy in the process of predicting a traditional Chinese medicine microscopic image, so that the aim of improving the accuracy of predicting the traditional Chinese medicine microscopic image is fulfilled. In order to solve the problems, the technical scheme adopted by the invention is as follows: a traditional Chinese medicine microscopic identification method based on a multidimensional channel attention mechanism comprises the following steps: Combining the characteristics of incomplete cell structure, stereo image and unbalanced characteristic distribution of the traditional Chinese medicine microscopic image in the Input layer to obtain a traditional Chinese medicine microscopic image data enhancement model, and enhancing the traditional Chinese medicine microscopic image data by utilizing the traditional Chinese medicine microscopic image data enhancement model; Fusing shallow characteristic information at a back bone layer, and fusing deep characteristic information befo