CN114882995B - A tongue image constitution identification method based on deep learning - Google Patents

A tongue image constitution identification method based on deep learning Download PDF

Info

Publication number
CN114882995B
CN114882995B CN202210608744.6A CN202210608744A CN114882995B CN 114882995 B CN114882995 B CN 114882995B CN 202210608744 A CN202210608744 A CN 202210608744A CN 114882995 B CN114882995 B CN 114882995B
Authority
CN
China
Prior art keywords
tongue
model
inception
image
enhancement
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202210608744.6A
Other languages
Chinese (zh)
Other versions
CN114882995A (en
Inventor
李聪聪
闫新晟
滕桂法
王芳
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Heibei Agricultural University
Original Assignee
Heibei Agricultural University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Heibei Agricultural University filed Critical Heibei Agricultural University
Priority to CN202210608744.6A priority Critical patent/CN114882995B/en
Publication of CN114882995A publication Critical patent/CN114882995A/en
Application granted granted Critical
Publication of CN114882995B publication Critical patent/CN114882995B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/243—Classification techniques relating to the number of classes
    • G06F18/2431—Multiple classes
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/70—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for mining of medical data, e.g. analysing previous cases of other patients
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Physics & Mathematics (AREA)
  • Medical Informatics (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Databases & Information Systems (AREA)
  • Biomedical Technology (AREA)
  • Public Health (AREA)
  • Multimedia (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computational Linguistics (AREA)
  • Mathematical Physics (AREA)
  • Molecular Biology (AREA)
  • Pathology (AREA)
  • Biophysics (AREA)
  • Epidemiology (AREA)
  • Primary Health Care (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Image Analysis (AREA)

Abstract

本发明所述一种基于深度学习的舌象体质辨别方法部署在移动终端，方便用户随时随地识别舌象图像。本发明方法采用卷积神经网络算法，对舌图像进行了同类增强和混类增强两种方法组合的方式对舌象数据集进行数据增强。使得模型能够适应各种环境下的舌象数据，提取的特征能够直接输出分类结果，计算量大大减少且泛化能力更强。

The method for distinguishing tongue constitution based on deep learning described in the present invention is deployed on a mobile terminal, which is convenient for users to identify tongue images anytime and anywhere. The method of the present invention adopts a convolutional neural network algorithm, and performs data enhancement on the tongue image data set by combining the same type enhancement and mixed type enhancement methods. This enables the model to adapt to tongue image data in various environments, and the extracted features can directly output the classification results, which greatly reduces the amount of calculation and has stronger generalization ability.

Description

Tongue condition distinguishing method based on deep learning
Technical Field
The invention relates to the field of images, in particular to a tongue constitution distinguishing method based on deep learning.
Background
At present, image recognition by using artificial intelligence techniques such as machine learning, deep learning and the like tends to be mature. Some students build a tongue image database by collecting a large number of tongue images with physique labels, and design algorithms to match the tongue images with the built database, so as to achieve the aim of recognition. Considering the complexity of the traditional manual tongue feature extraction and the accuracy of tongue region segmentation, deep learning is applied to constitution recognition due to the advantages of automatic feature extraction, and some students adopt various models of convolutional neural networks to perform tongue constitution recognition. Although these studies have driven the development of tongue diagnosis intelligent physique identification, there are also some general problems, firstly, deep learning requires a large number of data sets as support, and sample data acquisition of tongue images has a certain difficulty. Secondly, the classification is simply carried out by using the model, the model is not improved and optimized, and the accuracy of the constitution identification by the method is not high.
Disclosure of Invention
In order to solve the problems, the application provides a tongue image physique identification method based on deep learning, in particular to a tongue image identification method based on an improved convolutional neural network GoogleNet model, which is used for physique identification based on an acquired tongue image data set, and the main contribution of the application is as follows: A tongue picture data set with constitution labels is constructed, and a GoogleNet network model (TCR-GoogleNet) suitable for tongue picture constitution identification is provided by adding batch normalization operation, improving and optimizing Inception module, adding attention mechanism, combining different Inception structures and the like. The ablation experiment is designed to verify the contribution degree of each improved module to the prediction precision, and the comparison experiment is designed with the mainstream network of 5 image classifications, so that the accuracy and the stability of the model provided by the application are improved greatly.
The embodiment of the invention provides the following technical scheme:
A tongue constitution distinguishing method based on deep learning is deployed at a mobile terminal and comprises the following steps:
acquiring a tongue image to be identified, and identifying the tongue image to be identified by using a trained improved convolutional neural network GoogleNet model to obtain an identification result of the tongue image;
Displaying the identification result of the tongue picture image on the mobile terminal;
wherein, the training of the improved convolutional neural network GoogleNet model is as follows:
The method comprises the steps of obtaining a tongue image data set, and carrying out enhancement processing on the tongue image data set to obtain a tongue image training data set, wherein the enhancement processing comprises two modes of similar enhancement and mixed enhancement;
Training the improved convolutional neural network GoogleNet model by using the tongue image training dataset to obtain a trained improved convolutional neural network GoogleNet model;
Wherein the improved convolutional neural network GoogleNet model improvement portion comprises:
Each middle layer of the model is subjected to standardized treatment, so that gradient disappearance and gradient explosion are effectively avoided, the network convergence speed is accelerated, and the network structure is optimized;
The residual errors are fused into Inception structures, the Inception structures are improved and optimized, and the input features and the features output by Inception structures are subjected to feature fusion, and the name is Inception B;
3x3 convolution branches in Inception modules are decomposed into 1x3 and 3x1 convolutions, 5x5 convolutions are decomposed into 3x3 convolutions and 1x3 and 3x1 convolutions, batch Normalization is added after each branch, and parameter quantity of a model is effectively reduced and named Inception C;
On the basis of GoogleNet original models, inception modules which are not improved are reserved, the modules are named as InceptionA, inceptionA, inceptionB and Inception C structures are fused and applied to a GoogleNet network model after improvement, inceptionA modules are added in a shallow layer, inceptionB modules are added in a middle layer, inception C modules are added in a deep layer, wherein the A and B modules are respectively used twice, and the C module has stronger extraction and perception capability on characteristics, so that the method is used for five times.
The similar enhancement means that the image of the same label is subjected to data normalization, horizontal or vertical overturn, image scaling up or down, random rotation, random miscut transformation, brightness change and pixel filling method to expand the data set, and the color is also an important characteristic in tongue condition identification, so that the tongue condition data set is not transformed in terms of color channel.
The method for mixing type enhancement aims at solving the problem of unbalance among sample data sets, the method for mixing type enhancement mixes sample data of two types of different labels to construct a virtual training sample rich tongue image data set, and the mathematical expression is shown in (1):
Wherein (x i,yi)(xj,yj) is two samples and corresponding labels selected randomly, lambda is a random number obeying beta distribution, and lambda epsilon [0,1].
Compared with the prior art, the technical scheme has the following advantages:
The tongue picture constitution distinguishing method based on deep learning is deployed in the mobile terminal, so that a user can conveniently recognize tongue picture images at any time and any place. The method adopts a convolutional neural network algorithm to carry out data enhancement on the tongue image data set in a mode of combining two methods of similar enhancement and mixed enhancement on the tongue image. The model can adapt to tongue image data in various environments, the extracted features can directly output classification results, the calculated amount is greatly reduced, and the generalization capability is stronger.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings that are required in the embodiments or the description of the prior art will be briefly described, and it is obvious that the drawings in the following description are some embodiments of the present invention, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 is a schematic flow chart of a tongue constitution distinguishing method based on deep learning;
FIG. 2 is a proposed method architecture;
FIG. 3 is a sample image;
FIG. 4 is a data enhanced image;
FIG. 5 is Mixup visualizations;
FIG. 6 is a Inception B structure;
FIG. 7 is a Inception C structure;
FIG. 8 is a SE module;
FIG. 9 is a InceptionA structure;
FIG. 10 is a diagram showing a structure of a tongue constitution identification model;
FIG. 11 is an accuracy curve (1) and a loss value curve (2) during training;
FIG. 12 is a graph of accuracy training for four model test sets.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the application. Thus, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
As shown in fig. 1, the application provides a tongue constitution distinguishing method based on deep learning, which comprises the following steps:
acquiring a tongue image to be identified, and identifying the tongue image to be identified by using a trained improved convolutional neural network GoogleNet model to obtain an identification result of the tongue image;
Displaying the identification result of the tongue picture image on the mobile terminal;
wherein, the training of the improved convolutional neural network GoogleNet model is as follows:
The method comprises the steps of obtaining a tongue image data set, and carrying out enhancement processing on the tongue image data set to obtain a tongue image training data set, wherein the enhancement processing comprises two modes of similar enhancement and mixed enhancement;
Training the improved convolutional neural network GoogleNet model by using the tongue image training dataset to obtain a trained improved convolutional neural network GoogleNet model;
Wherein the improved convolutional neural network GoogleNet model improvement portion comprises:
Each middle layer of the model is subjected to standardized treatment, so that gradient disappearance and gradient explosion are effectively avoided, the network convergence speed is accelerated, and the network structure is optimized;
The residual errors are fused into Inception structures, the Inception structures are improved and optimized, and the input features and the features output by Inception structures are subjected to feature fusion, and the name is Inception B;
3x3 convolution branches in Inception modules are decomposed into 1x3 and 3x1 convolutions, 5x5 convolutions are decomposed into 3x3 convolutions and 1x3 and 3x1 convolutions, batch Normalization is added after each branch, and parameter quantity of a model is effectively reduced and named Inception C;
On the basis of GoogleNet original models, inception modules which are not improved are reserved, the modules are named as InceptionA, inceptionA, inceptionB and Inception C structures are fused and applied to a GoogleNet network model after improvement, inceptionA modules are added in a shallow layer, inceptionB modules are added in a middle layer, inception C modules are added in a deep layer, wherein the A and B modules are respectively used twice, and the C module has stronger extraction and perception capability on characteristics, so that the method is used for five times.
The similar enhancement means that the image of the same label is subjected to data normalization, horizontal or vertical overturn, image scaling up or down, random rotation, random miscut transformation, brightness change and pixel filling method to expand the data set, and the color is also an important characteristic in tongue condition identification, so that the tongue condition data set is not transformed in terms of color channel.
The method for mixing type enhancement aims at solving the problem of unbalance among sample data sets, the method for mixing type enhancement mixes sample data of two types of different labels to construct a virtual training sample rich tongue image data set, and the mathematical expression is shown in (1):
Wherein (x i,yi)(xj,yj) is two samples and corresponding labels selected randomly, lambda is a random number obeying beta distribution, and lambda epsilon [0,1].
The application has the following technical effects:
1. Various improvement modules are added in the original GoogleNet model, for example, data enhancement is carried out on the tongue image data set by adding a mode of combining two methods of similar enhancement and mixup mixed enhancement, batch normalization operation is added, and an attention mechanism module is added.
2. The base module Inception structure of GoogleNet is improved and optimized, for example, a residual structure is added into an original Inception structure for feature fusion, convolution substitution and decomposition operations are added into a Inception structure, and the improved two Inception structures and the original Inception structure are combined together to construct a model.
3. The method adopts a convolutional neural network algorithm to carry out data enhancement on the tongue image data set in a mode of combining two methods of similar enhancement and mixed enhancement on the tongue image. The model can adapt to tongue image data in various environments, the extracted features can directly output classification results, the calculated amount is greatly reduced, and the generalization capability is stronger.
The method is improved and optimized based on the GoogleNet model of the current stage main stream, and the classifier adopts a softmax classifier, so that the method is suitable for identifying multi-classification problems such as physique by a supervised learning method, and better effects are obtained on a calculation method and an identification effect.
Compared with the traditional Chinese medicine constitution identification judging time, the diagnosis time is reduced, the diagnosis efficiency is improved, and compared with the machine learning deep learning method at the present stage, the identification accuracy is improved.
The method of the invention is based on a large number of tongue image data sets, applies the deep learning technology to the traditional Chinese medicine physique recognition field, can not only judge physique through the PC end, but also judge physique through transplanting to the mobile phone end, and is very convenient, high in accuracy and time-saving.
The method combines the deep learning method with the traditional Chinese medicine physique identification, identifies on the basis of big data, effectively solves the problem of difficult identification of the traditional physique, and has certain market value and popularization value.
The specific method for distinguishing tongue constitution based on deep learning is described below.
1. Introduction to the invention
Traditional Chinese medicine (Traditional CHINESE MEDICINE, TCM) is a treasure of China, has thousands of years of history of disease prevention and treatment, plays an irreplaceable role in medical care and disease treatment of China, and is accepted and frequently used by more and more people internationally. The simultaneous treatment of different diseases and different diseases is the basic principle of understanding and treating diseases in traditional Chinese medicine, and the premise and basis of the treatment method based on syndrome differentiation depend on the constitution of individuals. In the theory of traditional Chinese medicine, constitution is considered to be a relatively stable inherent characteristic which is expressed by a human body on the basis of congenital conditions and acquired, and comprises the morphological structure and psychological qi quality of the human body. The constitution of traditional Chinese medicine is closely related to the health of human body, and the constitution is the basis and core content of the constitution research of traditional Chinese medicine, which influences the diagnosis and treatment of the whole clinical discipline of traditional Chinese medicine. In combination with the definition of constitutions and the clinical application principle of modern constitutions proposed in Huangdi's interior channel, wang Qi teaches a constitution nine-division method, namely, the constitution type is classified into mild constitutions, qi deficiency constitutions, yang deficiency constitutions, yin deficiency constitutions, phlegm-dampness constitutions, damp heat mass, blood stasis constitutions, qi depression constitutions and specific endowment constitutions, wherein the mild constitutions are healthy constitutions, and the other eight constitutions are deviant constitutions.
The tongue diagnosis is the most commonly used means in the diagnosis of traditional Chinese medicine, which considers that tongue manifestations are related to the veins and viscera of the human body and reflect the internal environment of the body, such as cold, summer-heat, deficiency and excess, and yin-yang excess and excess. The doctor judges the health condition of the patient and the susceptibility to the diseases by observing the changes of the color, the texture and the morphology of the tongue fur, and gives reasonable advice. Clinical practice also proves that tongue diagnosis is a simple, effective and without side effects, and that tongue manifestations contain information directly related to the constitution of the human body, so that it is feasible to use tongue diagnosis for constitution identification. In 2009, "classification and determination of constitutions in Chinese medicine" describes tongue features corresponding to each constitutions in detail, for example, "the tongue is fat and tender, has teeth marks on the edge, and the tongue is pale or slightly pale and dark, and the tongue coating is smooth" is the tongue features of people with constitutions due to yang deficiency, besides, other physical features of constitutions are introduced, and these physical features can be used for designing questionnaires for identifying constitutions. However, the questionnaire has great subjectivity, the individual can hardly make objective selection, the problem of the scale is more, a great deal of time is required to be consumed, and meanwhile, the judgment accuracy is also affected. With the popularization of artificial intelligence methods such as machine learning, deep learning and the like, if tongue images of individuals can be shot and stored in a digital image form, intelligent traditional Chinese medicine physique identification based on tongue diagnosis can be realized by combining an artificial intelligence technology, so that more objective and accurate results can be obtained.
At present, image recognition by using artificial intelligence techniques such as machine learning, deep learning and the like tends to be mature. Some students build a tongue image database by collecting a large number of tongue images with physique labels, and design algorithms to match the tongue images with the built database, so as to achieve the aim of recognition. The scholars also design a quantification method to extract objective tongue picture characteristics such as chromaticity and saturation of tongue and tongue fur, texture characteristics, smoothness and tooth trace characteristics of the tongue fur and the like, and establish a traditional Chinese medicine physique identification model based on machine learning methods such as a Support Vector Machine (SVM), K-means, an artificial neural network and the like. Considering the complexity of the traditional manual tongue feature extraction and the accuracy of tongue region segmentation, deep learning is applied to constitution recognition due to the advantages of automatic feature extraction, and some students adopt various models of convolutional neural networks to perform tongue constitution recognition. Although these studies have driven the development of tongue diagnosis intelligent physique identification, there are also some general problems, firstly, deep learning requires a large number of data sets as support, and sample data acquisition of tongue images has a certain difficulty. Secondly, the classification is simply carried out by using the model, the model is not improved and optimized, and the accuracy of the constitution identification by the method is not high. In order to solve the problems, the application provides a new method based on an improved convolutional neural network GoogleNet model, and the method is used for body constitution identification based on an acquired tongue image dataset, and the main contribution of the application is that a tongue image dataset with body constitution labels is constructed, and a GoogleNet network model (TCR-GoogleNet) suitable for tongue image body constitution identification is provided by adding a batch normalization operation, an improved optimization Inception module, an attention adding mechanism, combined use of different Inception structures and the like. The ablation experiment is designed to verify the contribution degree of each improved module to the prediction precision, and the comparison experiment is designed with the mainstream network of 5 image classifications, so that the accuracy and the stability of the model provided by the application are improved greatly.
2. Method of
This section describes a specific process of physique recognition by tongue images. The first section briefly describes the specific steps and architecture of the method. The second section briefly illustrates the construction of a tongue image set, including the basic flow and requirements of image acquisition and the method of image preprocessing. The third section describes preprocessing of tongue images and data enhancement operations on tongue image datasets. The fourth section describes an improved method of module and construction of the model as a whole.
2.1 Method architecture
In order to better identify the constitution of the tongue, the application provides a TCR-Googlenet model, and a good constitution identification result is obtained by training a large amount of data in the constructed tongue data set. The infrastructure of this method is shown in fig. 2, with a total of five steps. The first step is the acquisition of tongue images. The second step and the third step are image preprocessing, which comprises cutting out face images to extract tongue areas, and the quality of tongue image data sets can be greatly improved after cutting out. And the fourth step is to input the extracted tongue image into the established deep neural network model for training and constitution recognition. And finally classifying and identifying different physique types.
2.2 Dataset construction
(1) Description of the characteristics of nine common constitutions
In the tongue diagnosis of traditional Chinese medicine, different constitutions have different tongue appearance characteristics, and doctors often judge constitutions and pathology of patients through the different tongue appearance characteristics. According to the theory [6] proposed by Wang Qi professor, the tongue characteristic description of nine constitutions of peace quality, qi deficiency quality, yang deficiency quality, yin deficiency quality, phlegm dampness quality, damp heat mass, blood stasis quality, qi depression quality and specific endowment quality is taken as a research object, and the tongue characteristic description is shown in table 1, and can provide a visual basis for subjective discrimination of the deep learning network model on the accuracy of Chinese medicine constitution identification.
Table 1 tongue characteristic description of nine constitutions
(2) Tongue constitution data set construction
The deep learning network model simulates the condition that the human brain abstract extracts image features from low level to high level based on a large-scale image dataset, the tongue image dataset needs to be constructed for tongue image physique recognition based on deep learning, tongue images of patients are acquired in a clinic in a middle hospital, 1752 tongue image data are obtained in total, the tongue images are acquired, meanwhile, the types of physique are determined by filling in a traditional Chinese medicine physique questionnaire by the patients, tongue image tag marking is carried out, tongue region images are accurately extracted from each image through manual processing, and therefore the influence on tongue image physique recognition of other parts in the originally acquired images is reduced. Tongue images are taken under brightly lit natural conditions. The sample distribution of the tongue image dataset is shown in table 2, where the numerical value represents the number of tongue images per constitution. The corresponding samples for each constitution are shown in fig. 3.
Table 2 data set sample distribution
2.3 Data enhancement
Because the images shot by the mobile phone camera are affected by external factors such as light rays, angles and the like, and the number of collected tongue images is relatively small, in order to prevent the phenomena of over fitting, low model robustness and the like, the collected images need to be subjected to data enhancement. The application mainly adopts two modes of similar enhancement and mixed enhancement to enhance the data set.
(1) Similar reinforcement
The same type of enhancement mainly refers to data set expansion of images of the same label by using methods such as data normalization, horizontal or vertical overturn, image scaling up or down according to proportion, random rotation, random miscut transformation, brightness change, pixel filling and the like, and color is also an important characteristic in tongue condition identification, so that the tongue condition data set is not transformed in terms of color channels. Fig. 4 is a picture of a data-like enhancement.
(2) Mixed type enhancement
The method for mixing type enhancement aims at solving the problem of unbalance among sample data sets, for example, the specific nature is an unusual physique type, so that sample data are relatively few. The mathematical expression is shown as (1):
wherein (x i,yi)(xj,yj) is two samples and corresponding labels selected randomly, lambda is a random number obeying beta distribution, and lambda epsilon [0,1]. Fig. 5 is an effect graph of Mixup randomly generated twice.
2.4 Construction of tongue image physique recognition model
The application improves GoogleNet models for tongue image recognition, googleNet is a winner of 2014 ILSVRC contests, and belongs to deep convolutional neural network models. Since the development of VGG models, convolutional neural networks have evolved towards deeper architectures, including more and more network layers and parameters, however, adding networks together necessarily places a burden on computing resources. The Inception module provided in GoogleNet effectively relieves the problems of gradient elimination and gradient explosion caused by deepening a network, improves the operation efficiency of the model, and provides a model of a deep neural network based on the GoogleNet model after improvement. This section mainly introduces the construction of tongue image recognition models and the improvement strategy.
(1) Batch standardization operation (BatchNormalization, BN)
Along with the continuous updating of parameters in the neural network, the data distribution input by each layer in the middle often has larger difference from the data distribution before the updating of the parameters, so that the network needs to be continuously adapted to the new data distribution, and the difficulty of model training can be increased. BN operation proposed in 2015 effectively solves this problem. The application performs standardization treatment on each middle layer (before activating the function) of the model, effectively avoids the problems of gradient disappearance, gradient explosion and the like, accelerates the network convergence speed and optimizes the network structure. The calculation flow of BN is as follows:
If a batch of input variables is x 1,x2…,xk, the mean and variance are calculated respectively, and the mathematical expression is shown in (2) (3):
then, normalizing the input batch of x values to a mean value of 0 and a variance of 1, wherein the mathematical expression is as shown in (4):
Equation (4) corrects x i from the edge where the gradient is vanished to the region where the gradient is large, increasing the training speed, where ε is a small value, in order to avoid that the variance denominator is 0 and has no effect on the calculation.
And finally reconstructing the normalized value, wherein the mathematical expression is shown as (5):
wherein alpha and beta are two variables which can be learned, and the purpose is to compensate the nonlinear expression capacity of the model and recover the distribution which is learned by the network.
(2) Inception structural residual design
He proposed a residual structure in 2015, and proved the advantage of feature additive combination through theory and practice. In image recognition, the residual structure can effectively avoid the network degradation problem caused by depth deepening, and meanwhile, the gradient problem is solved, so that the performance of the network is improved. The application refers to the idea that the residual error is fused into Inception structure, the Inception structure is improved and optimized, the input characteristic and the characteristic output by Inception structure are fused, and fig. 6 is a Inception module added with the residual error, named Inception B.
(3) Inception structural convolution decomposition and substitution
The decomposition and replacement of convolutions is presented in Inceptionv as decomposing a large convolution kernel into symmetrical small convolution kernels and into asymmetrical convolution kernels. The first method is to replace one convolution kernel of 5x5 with two convolution kernels of 3x 3. The second method is to replace the convolution kernel of n by 1*n and n1 and then stack. Both methods can reduce the number of parameters and increase more non-linear capacity with unchanged receptive field. Therefore, the application adopts the two methods to improve Inception modules, the 3x3 convolution branches in Inception modules are decomposed into 1x3 convolution and 3x1 convolution, the 5x5 convolution is decomposed into 3x3 convolution and 1x3 convolution and 3x1 convolution, batch Normalization is added after each branch, and the parameter number of the model is effectively reduced. Fig. 7 shows a Inception structure after modification by this method, designated Inception C.
(4) SE attention module
The SE module is composed of two operations, namely, squeeze and expression, wherein Squeeze is low-dimensional embedding of global information, is a compression process, expression is a mapping transformation process, and the SE module mainly explores the relation among channels and gives weight to network channels to enhance the learning ability of the network, namely, channel attention. Fig. 8 is a schematic diagram of a SE module.
(5) Inception structural fusion and model construction
The application retains the Inception module without improvement on the basis of GoogleNet original model, which is named Inception A as shown in figure 9. The Inception A and Inception B and Inception C structures proposed above are fused and applied to a GoogleNet network model after improvement, a InceptionA module is added in a shallow layer, a Inception B module is added in a middle layer, and a Inception C module is added in a deep layer, wherein the A module and the B module are respectively used twice, and the C module has stronger extraction and perception capabilities on characteristics, so that the method is used for five times.
In summary, the method and the improvement mode are provided, and the overall structure diagram of the tongue image physique identification model after improvement is shown in fig. 10.
3. Experimental design and results analysis
The experiment adopts a win10 operating system, the CPU is Intel (R) Core (TM) i9-7900X3.30GHz, the GPU is GeForce RTX 3090, python version is 3.8.4, cuda and Cudnn versions are 11.2 and 8.1.0 respectively, and a tensorflw 2.5.0 framework is used as an experimental environment.
3.1 Parameter selection and optimization
In the deep learning model training process, the accuracy of the model is often influenced by the selection of parameters, in the parameter selection and optimization process, the method adopts a repeated ablation experiment for fine adjustment, a Batch training mode is adopted when the model is used for training, the Batch size (Batch size) is 32, 100 rounds of Epoch are performed in total, an Adam algorithm (Adaptive Moment Estimation adaptive moment estimation of self-adaptive momentum) is adopted as an optimization function, the momentum is 0.9, the weight attenuation value is 1 multiplied by 10 -5, and the initial learning rate is 0.0001. To further prevent overfitting, the Dropout value is set to 0.3, i.e., 30% of the nodes are randomly discarded, and the classifier uses SoftMax.
3.2 Evaluation index
In order to evaluate the advantages and disadvantages of the model, the number of rounds (E) of model convergence, namely the convergence speed, the confusion matrix, the loss value (L) and the Top-1 accuracy (Acc), are selected as evaluation indexes in model training, wherein the number of rounds required by model convergence means the evaluation of the model convergence speed under the same condition, the physique identification is a multi-classification problem, the loss function adopts a multi-classification cross entropy loss function (cross entropy loss), and the loss value expression is as follows:
Wherein y i is a real label corresponding to the ith sample, p i is a predicted value of the model training image, N is the total number of physique categories, and K is the total number of tongue picture samples.
The Top-1 accuracy reflects the proportion of the correct images to all the recognition images in the recognition result, reflects the training effect of the model on the data set, and has the mathematical expression:
The model prediction method comprises the steps of TP (True positive), TN (True negative), FN (False negative) and FP (False positive), wherein the model prediction method is positive samples, TN (True negative) is negative samples of the model prediction method, and 5326 is negative samples of the model prediction method.
3.3 Ablation experiments
The ablation experiment is used as a common method for model evaluation, namely the effectiveness of an improved algorithm is further explained by deleting a certain part of a network model so as to facilitate better understanding of the influence and contribution degree of the improved network module on the accuracy of the network model.
The experiment is combined and designed from the aspects of adding batch standardization operation, respectively using and combining Inception A, inceptionB and Inception C structures and adding attention mechanism, 7 groups of experiments are designed, the serial numbers are 1-7, and the experimental design is shown in table 3.
Table 3 ablation experimental design
Table 4 comparison of ablation experimental results
The experimental results are shown in table 4, and the following conclusions can be drawn from each group of experimental results:
(1) Batch normalization operation (BN)
The model is GoogleNet original model only by Inception A modules in experiment 1, BN operation is added in experiment 2, and compared with the result, the accuracy of experiment 2 is improved by 1.6% compared with the Top-1 of experiment 1, which shows that the BN operation optimizes the network structure and improves the model accuracy to a certain extent.
(2)Mixup
The Mixup mixed enhancement method is used in the experiment 3 to expand the data set, the accuracy and the loss value of the comparison experiment 2 are optimized to different degrees, the defect of less tongue image data is made up to a certain degree by combining the method with the similar data enhancement, and the model is better fitted with the data set while almost no additional calculation cost exists.
(3) Inception structure
The Inception A modules in the model are replaced by Inception B modules added with residual structures in the experiment 4, but the accuracy of the model is reduced by about 2.5%, and the widening of the network by the stacked residual structures with one taste is proved not to be very suitable for the dataset of the application.
The Inception C modules are all used in experiment 5, so that the Top-1 accuracy is improved obviously, and the fact that the network structure deepened through convolution decomposition and replacement has a better effect on extracting tongue image characteristic information is explained, and the recognition accuracy is improved while the parameters of a network model are reduced.
Experiment 6 combines the operations of experiment 2, experiment 4 and experiment 5, adopts three Inception structure fusion modes to improve the model, widens and deepens the network model, extracts tongue image characteristics more comprehensively, and has better accuracy and loss value compared with the independent Inception structure effect.
(4) SE module
Experiment 7 is the final model after the improvement of the study, and as can be seen from the table, compared with the previous experiment, the accuracy of the Top-1 of the model can be improved by 1.3% by adding the SE module, so that the attention-introducing mechanism can improve the accuracy of the Top-1 of the model without increasing the depth and width of the network and obviously increasing the network parameters.
3.4 Experimental results and analysis
Based on the construction of the model, 85% of the data set is used as a training set, 15% is used as a test set, random reproduction of data is ensured when the data set is divided, and the experimental effect of the model is analyzed from three aspects of confusion matrix, the model training set, the test set index and model comparison.
(1) Confusion matrix
The present model was applied to a test set of tongue image data sets and the resulting confusion matrix is shown in table 5. Which includes the number of samples of the 9 physical images in the test set that were correctly predicted and incorrectly predicted. In the confusion matrix, the values on the main diagonal are the number of samples predicted correctly, and the values at other positions are the number of samples predicted incorrectly.
Table 5 confusion matrix for test set
It can be known from the table that qi deficiency and yang deficiency are easily confused in the classification process, because the two constitutions may have closer tongue characteristics and smaller differences between the types, so that the model easily confuses the two constitutions. The tongue features of blood stasis and qi stagnation are obviously different from those of other constitutions, and the classification effect is good. The recognition rate of the specific constitution is the lowest, because the samples of the specific constitution are few, the number of people with the specific constitution is too few in reality, a large number of training samples are difficult to collect, and tongue image characteristics of the specific constitution are varied. The recognition accuracy of phlegm-dampness and damp-heat is highest, and the model can fully extract tongue image characteristics due to the fact that the training samples are sufficient.
(2) Training set and test set recognition results
Fig. 11 shows an accuracy curve and a loss value curve in the training process of the training set and the test set, where the training accuracy curve (1) can reflect the variation trend of the model prediction accuracy with the increase of the number of epochs. The training loss value curve (2) shows the deviation and fluctuation condition between the model predicted value and the real value along with the increase of the number of the Epoch rounds, and the lower the loss value, the higher the prediction precision of the model is, and the lower the probability of model prediction error is indicated. The accuracy of the training set of the model reaches 92.2%, the loss value is 0.375, the accuracy of the test set reaches 87.4%, and the loss value is 0.501. Compared with other researches of the same type at the present stage, the model has higher recognition accuracy and stability.
(3) Model comparison results
In order to verify the effectiveness of the model and the effect after improvement, five main stream convolutional neural network models of image recognition are selected for comparison experiments, and FIG. 12 is an accuracy line diagram comparison of the model of the application and the VGG-16, resnet-50, mobilenet-v3-large, mobilenet-v3-small, googleNet original model for test set training, so as to show the difference between the model and other models. Table 3 visually lists the Top-1 accuracy, loss values, and number of epochs that reached convergence (convergence rate) for the 6 model test sets. The accuracy of the test set is regarded as a standard for model comparison, and the test set can better represent the fitting capacity of the model to the data set, so that the training accuracy of the model is accurately evaluated. In model evaluation, the priority of model prediction accuracy is greater than the convergence speed.
In the experiment and the comparison experiment, the neural network model is trained from the head, so that the initial accuracy is lower, but the model can reach the highest Top-1 accuracy at a higher speed, which is 87.4%. Although the model of the application has slower convergence rate than the model GoogleNet, mobilenet-v3-large, mobilenet-v3-small, the Top-1 accuracy of the model is 11.5% higher than that of the model GoogleNet, 17.1% higher than that of the model mobilenet-v3-large and 19.5% higher than that of the model mobilenet-v 3-small. The mobilenet-v3 model has the advantages that the depth separable convolution is adopted, so that the operand and the parameter quantity are greatly reduced, the mobilenet-v3 series has very good advantages in the aspect of convergence speed, but the oscillation is larger after the convergence on the data set, the stability of the model is poor, and the accuracy is low. The Resnet-50 network model added with the residual error has better performance on the Top-1 accuracy and the loss value, which indicates that the fused residual error structure has better characterization capability on tongue image characteristics. Therefore, a residual structure is added in the model construction process, and then other improvement modules are combined, so that the recognition effect is improved by 5.6% compared with the Top-1 accuracy of Resnet-50, the loss value is reduced by 0.112, and the convergence rate is also improved slightly. The VGG-16 model improves model training precision by deepening the network layer number and reducing the convolution kernel size, but the effect in a comparison experiment does not reach the expected value, the accuracy is only 72.8%, and a large gap is reserved between the model training accuracy and the model training result of the application, so that the tongue picture characteristics can not be extracted well by a simple deepened network structure. In the network model, the model is improved from the depth and the width of the network, and the model prediction accuracy is improved to a certain extent by a plurality of improved modules added on the basis of the model improvement.
Table 6 comparison of model results
In a comprehensive view, the improved model of the application has larger improvement on recognition accuracy than other models, and has moderate convergence rate, but the priority of model prediction accuracy is higher than the convergence rate, so the recognition effect of the improved model on the tongue image dataset is superior to that of other 5 models.
The application provides a TCR-GoogleNet model based on GoogleNet, a mature tongue image data set is constructed, and meanwhile, the image is continuously collected, so that the tongue image data set is enlarged, and more powerful data support is provided for subsequent research work. The model shows stronger feature extraction capability and better recognition accuracy on tongue image data sets under the advantages of the width and depth of a network and various improvements. The method aims at the nine traditional Chinese medicine constitutions of mild constitution, qi deficiency constitution, yang deficiency constitution, yin deficiency constitution, phlegm dampness constitution, damp heat constitution, blood stasis constitution, qi depression constitution and specific endowment constitution, so that the difficulty that abundant expert experience is needed for constitution classification is relieved to a certain extent, the constitution identification time is greatly shortened compared with a questionnaire manner, and a reference and thinking are provided for the objectification process of tongue diagnosis.
The tongue diagnosis-based constitution identification has a larger development space in the future, the tongue image data set can be more abundant and expanded, and the identification accuracy on the basis can be further improved. Since each person's constitution has a variety, one person may contain a variety of constitutions, so subsequent studies can be developed around multi-tag learning of tongue image constitution recognition. Besides, the study of tongue diagnosis can be used for identifying constitution, and the combination of tongue diagnosis and western medicine has been studied for preventing and treating diseases, but the study is still to be further developed.
In the present description, each part is described in a progressive manner, and each part is mainly described as different from other parts, and identical and similar parts between the parts are mutually referred.
The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the application. Thus, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (3)

1. A tongue constitution distinguishing method based on deep learning is characterized by being deployed in a mobile terminal and comprising the following steps:
acquiring a tongue image to be identified, and identifying the tongue image to be identified by using a trained improved convolutional neural network GoogleNet model to obtain an identification result of the tongue image;
Displaying the identification result of the tongue picture image on the mobile terminal;
the training process of the improved convolutional neural network GoogleNet model is as follows:
The method comprises the steps of obtaining a tongue image data set, and carrying out enhancement processing on the tongue image data set to obtain a tongue image training data set, wherein the enhancement processing comprises two modes of similar enhancement and mixed enhancement;
Training the improved convolutional neural network GoogleNet model by using the tongue image training dataset to obtain a trained improved convolutional neural network GoogleNet model;
Wherein the improved convolutional neural network GoogleNet model improvement portion comprises:
Each middle layer of the model is subjected to standardized treatment, so that gradient disappearance and gradient explosion are effectively avoided, the network convergence speed is accelerated, and the network structure is optimized;
The residual errors are fused into Inception structures, the Inception structures are improved and optimized, and the input features and the features output by Inception structures are subjected to feature fusion, and the name is Inception B;
3x3 convolution branches in Inception modules are decomposed into 1x3 and 3x1 convolutions, 5x5 convolutions are decomposed into 3x3 convolutions and 1x3 and 3x1 convolutions, batch Normalization is added after each branch, and parameter quantity of a model is effectively reduced and named Inception C;
On the basis of GoogleNet original model, a Inception module which is not improved is reserved, the model is named Inception A, inception A, inception B and Inception C structures are fused and applied to the GoogleNet network model after improvement, inception A modules are added in a shallow layer, inception B modules are added in a middle layer, inception C modules are added in a deep layer, wherein the A and B modules are respectively used for two times, and the C module has stronger extraction and perception capability on characteristics, so that the model is used for five times.
2. The method of claim 1, wherein the similar enhancement means that the image of the same label is subjected to data normalization, horizontal or vertical inversion, scaling up or down of the image, random rotation, random miscut transformation, brightness variation and pixel filling method, and the color is also an important feature in tongue condition identification, so that the tongue condition dataset is not transformed in terms of color channel.
3. The method of claim 2, wherein the mixing enhancement method is to solve the problem of unbalance between classes of sample data sets, and the mixing enhancement method mixes sample data of two different classes of labels to construct a virtual training sample rich tongue image data set, and the mathematical expression is as shown in (1):
Wherein (x i,yi)(xj,yj) is two samples and corresponding labels selected randomly, lambda is a random number obeying beta distribution, and lambda epsilon [0,1].
CN202210608744.6A 2022-05-31 2022-05-31 A tongue image constitution identification method based on deep learning Active CN114882995B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210608744.6A CN114882995B (en) 2022-05-31 2022-05-31 A tongue image constitution identification method based on deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210608744.6A CN114882995B (en) 2022-05-31 2022-05-31 A tongue image constitution identification method based on deep learning

Publications (2)

Publication Number Publication Date
CN114882995A CN114882995A (en) 2022-08-09
CN114882995B true CN114882995B (en) 2024-12-10

Family

ID=82679075

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210608744.6A Active CN114882995B (en) 2022-05-31 2022-05-31 A tongue image constitution identification method based on deep learning

Country Status (1)

Country Link
CN (1) CN114882995B (en)

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116228611A (en) * 2022-09-07 2023-06-06 南京邮电大学 Infectious disease image-oriented data enhancement method and system
CN116153494A (en) * 2022-11-17 2023-05-23 江苏省中医院 Model construction method, device and image system for assisting diagnosis and treatment of oral ulcer
CN116189856B (en) * 2022-11-23 2026-01-30 厦门大学 A lightweight multi-classification method and system based on tongue image features
CN116523808A (en) * 2023-05-29 2023-08-01 重庆电子工程职业学院 Image mixed data enhancement method capable of adaptively adjusting color information
CN119478579B (en) * 2024-10-10 2025-10-10 电子科技大学 A tongue image data enhancement method based on fractal features

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110189305A (en) * 2019-05-14 2019-08-30 上海大学 A multi-task tongue image automatic analysis method
AU2020103613A4 (en) * 2020-11-23 2021-02-04 Agricultural Information and Rural Economic Research Institute of Sichuan Academy of Agricultural Sciences Cnn and transfer learning based disease intelligent identification method and system

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11157814B2 (en) * 2016-11-15 2021-10-26 Google Llc Efficient convolutional neural networks and techniques to reduce associated computational costs
CN110570431A (en) * 2019-09-18 2019-12-13 东北大学 A Medical Image Segmentation Method Based on Improved Convolutional Neural Network
CN113989563A (en) * 2021-10-29 2022-01-28 河南科技大学 Multi-scale multi-label fusion Chinese medicine tongue picture classification method

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110189305A (en) * 2019-05-14 2019-08-30 上海大学 A multi-task tongue image automatic analysis method
AU2020103613A4 (en) * 2020-11-23 2021-02-04 Agricultural Information and Rural Economic Research Institute of Sichuan Academy of Agricultural Sciences Cnn and transfer learning based disease intelligent identification method and system

Also Published As

Publication number Publication date
CN114882995A (en) 2022-08-09

Similar Documents

Publication Publication Date Title
CN114882995B (en) A tongue image constitution identification method based on deep learning
Kwasigroch et al. Deep CNN based decision support system for detection and assessing the stage of diabetic retinopathy
CN119323574B (en) An AI-assisted decision support system for skin disease diagnosis
CN113689954A (en) Hypertension risk prediction method, device, equipment and medium
CN115410046A (en) Skin disease tongue picture classification model based on deep learning, establishing method and application
CN107799165A (en) A kind of psychological assessment method based on virtual reality technology
Xie et al. PIF-Net: A parallel interweave fusion network for knee joint segmentation
CN120319499A (en) A method and system for constructing a TCM syndrome library based on convolutional neural network
CN111341437A (en) Auxiliary system for judging digestive tract diseases based on tongue images
He et al. EnhanceDeepIris model for iris recognition applications
CN113130066A (en) Tongue diagnosis image identification method based on artificial intelligence
CN110008925A (en) An automatic skin detection method based on ensemble learning
CN118116576B (en) Intelligent case analysis method and system based on deep learning
CN120147270A (en) A diabetic retinopathy grading method based on a visual-linguistic pre-training model and a ranking-aware prompting strategy
Haihong et al. SAE-wAMD: A Self-Attention Enhanced convolution neural network for fine-grained classification of wet Age-Related Macular Degeneration using OCT Images
Zhu et al. RAMST-CNN: a residual and multiscale spatio-temporal convolution neural network for personal identification with EEG
CN121215232A (en) A Multimodal TCM Method for Differentiating Insomnia Based on Multi-Task Deep Learning
Vijayalakshmi et al. Development of prognosis tool for type-II diabetics using tongue image analysis
CN113724857A (en) Automatic diagnosis device for eye ground disease based on eye ground image retina blood vessel
CN113974627A (en) Emotion recognition method based on brain-computer generated confrontation
Elzennary et al. Early deep detection for diabetic retinopathy
CN118335292A (en) An interactive assistance system for TCM prescriptions for specific diseases
Wang et al. Glaucoma Fundus Image Recognition Based on Optimal ROI Proportion and VIT-B/16
Zhou et al. Enhancing tiny tissues segmentation via self-distillation
Hao et al. A Framework on Imaging and Phenotype Information Hierarchical Graph Convolutional Networks for Identification of Autism Spectrum Disorder

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant