WO2023115682A1 - 自适应随机块卷积核网络的高光谱中药材鉴别方法 - Google Patents

自适应随机块卷积核网络的高光谱中药材鉴别方法 Download PDF

Info

Publication number
WO2023115682A1
WO2023115682A1 PCT/CN2022/076024 CN2022076024W WO2023115682A1 WO 2023115682 A1 WO2023115682 A1 WO 2023115682A1 CN 2022076024 W CN2022076024 W CN 2022076024W WO 2023115682 A1 WO2023115682 A1 WO 2023115682A1
Authority
WO
WIPO (PCT)
Prior art keywords
band
hyperspectral
medicinal materials
chinese herbal
chinese medicinal
Prior art date
Application number
PCT/CN2022/076024
Other languages
English (en)
French (fr)
Inventor
毛建旭
尹阿婷
王耀南
张辉
刘彩苹
朱青
刘敏
曾凯
陈煜嵘
李亚萍
赵禀睿
苏学叁
Original Assignee
湖南大学
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 湖南大学 filed Critical 湖南大学
Publication of WO2023115682A1 publication Critical patent/WO2023115682A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G06N20/10—Machine learning using kernel methods, e.g. support vector machines [SVM]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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; CALCULATING OR 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

Definitions

  • the invention belongs to the field of medical hyperspectral intelligent detection and analysis, in particular to a method for identifying hyperspectral Chinese medicinal materials with an adaptive random block convolution kernel network.
  • the processing flow first uses methods such as noise reduction and scatter correction to preprocess the acquired hyperspectral images; then principal component analysis is used.
  • PCA principal component analysis
  • PLS-DA partial least squares discriminant analysis
  • this kind of processing flow cannot accurately establish an identification model, and the model is not universal, and the amount of hyperspectral data is very large, it is difficult to extract effective characteristic information of Chinese herbal medicines, and the identification accuracy of the tested Chinese herbal medicines is not high.
  • To identify the difficult problem of diverse and complex Chinese herbal medicines develop a rapid non-destructive detection method that can effectively extract the space-spectral feature information of Chinese herbal medicines and is applicable to all kinds of Chinese herbal medicines.
  • the present invention provides a method for identifying hyperspectral Chinese herbal medicines using an adaptive random block convolution kernel network.
  • a hyperspectral Chinese medicinal material identification method of an adaptive random block convolution kernel network includes the following steps:
  • Step S100 taking hyperspectral images of Chinese herbal medicines, and constructing a hyperspectral raw data set of Chinese herbal medicines;
  • Step S200 Using the optimal clustering framework to obtain the optimal band subset of the hyperspectral data set of Chinese herbal medicines, select the best characteristic bands of the hyperspectral data set of Chinese medicinal materials from the optimal band subset based on the cluster sorting strategy, and form the best Band characteristic image;
  • Step S300 using principal component analysis to reduce the dimensionality of the data in the original hyperspectral data set of Chinese medicinal materials, and using a random projection method to obtain random blocks from the reduced dimensionality hyperspectral data of Chinese medicinal materials as convolution kernels;
  • Step S400 modifying the convolution kernel with a pixel adaptive method to obtain an adaptive random block convolution kernel
  • Step S500 using a layered network to extract features of Chinese herbal medicines by using an adaptive random block convolution kernel and the best band feature image convolution;
  • Step S600 Combining the features of Chinese medicinal materials extracted by the layered network and the image data of the best band features to construct a hyperspectral training set and test set of Chinese medicinal materials;
  • Step S700 use SVM to train the training set to obtain a classification prediction model, and predict the hyperspectral test set of Chinese medicinal materials based on the classification prediction model, so as to realize the identification and classification of Chinese medicinal materials.
  • step S100 includes:
  • Step S110 using a hyperspectral sorting instrument to acquire hyperspectral images of Chinese herbal medicines, and performing reflectance correction on the collected hyperspectral images of Chinese herbal medicines;
  • Step S120 use the corrected image as a sample of the hyperspectral data set of Chinese herbal medicines, and construct an original hyperspectral data set of Chinese herbal medicines.
  • step S200 includes:
  • Step S210 Calculating the local density and intra-cluster distance of each band of the hyperspectral data of Chinese medicinal materials, and normalizing the intra-cluster distance;
  • Step S220 weighted calculation of local density and intra-cluster distance to obtain the contribution value of each band of the Chinese herbal medicine hyperspectral image
  • Step S230 Divide the hyperspectral image of Chinese medicinal materials into a preset number of band subsets by the K-means++ clustering method, select the band with the largest contribution value of each band subset among the preset number of band subsets, and calculate the relationship between the band and the The similarity matrices of other band subsets are summed, and the value obtained by the summation is recorded as F, and F is minimized to obtain a preset number of optimal band subsets;
  • Step S240 Re-select the band with the largest contribution value in each optimal band subset to obtain the best feature band to form the best band feature image.
  • step S210 includes:
  • Step S211 Calculating the local density of each band of the hyperspectral data of Chinese medicinal materials, specifically:
  • D ij is the similarity matrix
  • i and j are the i and j bands of the hyperspectral data of Chinese medicinal materials respectively
  • d c is the cut-off distance of the area where each band is located;
  • Step S212 Calculate the intra-cluster distance of each band of the Chinese herbal medicine hyperspectral data, specifically:
  • D ij is the similarity matrix
  • i and j are the i and j bands of the hyperspectral data of Chinese medicinal materials respectively
  • the intra-cluster distance ⁇ max of the point k with the highest local density in the hyperspectral data of Chinese medicinal materials is:
  • Step S213 Normalize the intra-cluster distance ⁇ i , specifically:
  • ⁇ i ( ⁇ i - ⁇ min )./( ⁇ max - ⁇ min )
  • ⁇ i is the intra-cluster distance of each band
  • ⁇ min is the intra-cluster distance of the point with the smallest local density in the hyperspectral data of Chinese medicinal materials
  • ⁇ max is the intra-cluster distance of the point with the largest local density in the hyperspectral data of Chinese medicinal materials .
  • step S220 is specifically:
  • R i is the contribution value of the i-th band
  • ⁇ i is the local density of the i-th band
  • ⁇ i is the intra-cluster distance of the i-th band.
  • F in step S230 is specifically:
  • w pk is the similarity matrix between the band with the largest contribution value and other band subsets.
  • step S300 includes:
  • Step S310 Perform principal component analysis on the data in the original hyperspectral data set of Chinese herbal medicines for dimensionality reduction and whitening to obtain dimensionality-reduced data X p , where, N is the number of image pixels, and p is the number of principal components of the image;
  • Step S320 Use the random projection method to select M random blocks in the dimensionally reduced data as the convolution kernel P', where, P i ' is the ith random block convolution kernel, and w ⁇ w is the size of the convolution kernel.
  • step S400 includes:
  • Step S410 performing bilateral filtering on the best band feature image to obtain the filtered best band feature image
  • step S500 includes:
  • Step S510 set the number of layers of the layered network as n;
  • Step S520 Extract the features of the first layer of Chinese herbal medicines according to the adaptive random block convolution kernel and the best band feature image convolution;
  • Step S530 Repeat steps S300 and S400 for the characteristics of the first layer of Chinese medicinal materials to obtain the second layer of adaptive random block convolution kernel, according to the second layer of adaptive random block convolution kernel and the first layer of Chinese medicinal materials The features are extracted by convolution to obtain the features of the second layer of Chinese herbal medicines;
  • Step S540 Repeat step S530 until the features of the nth layer of Chinese medicinal materials are extracted.
  • the hyperspectral Chinese herbal medicine identification method based on the above-mentioned adaptive random block convolution kernel network first obtains the optimal band subset of the hyperspectral image of Chinese herbal medicine based on the optimal clustering framework, and then uses the cluster sorting method to effectively select from the optimal band subset Select the best feature band; then use the random projection method to use the random block extracted from the hyperspectral image of the Chinese herbal medicine as the convolution kernel; then use the pixel adaptive method to modify the convolution kernel, and perform feature extraction based on the feature band image of the Chinese herbal medicine; Thirdly, the layered network is used to extract the characteristics of Chinese herbal medicines, and combined with the hyperspectral image data of Chinese herbal medicines in the best band, the hyperspectral training set and test set of Chinese herbal medicines are constructed; finally, the SVM is used to train the training set to obtain a classification prediction model.
  • the model predicts the test set of Chinese herbal medicines, realizes the identification and classification of Chinese herbal medicines, greatly improves the identification accuracy of Chinese herbal medicines, solves the identification problem of various types of Chinese herbal medicines and complex components, and is applicable to the rapid and non-destructive identification of various Chinese herbal medicines .
  • Fig. 1 is the flow chart of the hyperspectral Chinese medicinal material identification method of the self-adaptive random block convolution kernel network provided by an embodiment of the present invention
  • Fig. 2 is a schematic diagram of some samples of a hyperspectral Chinese herbal medicine data set according to an embodiment of the present invention
  • Fig. 3 is a schematic structural framework diagram of an adaptive random block convolution kernel network model according to an embodiment of the present invention.
  • Fig. 4 is a schematic structural diagram of an adaptive random block convolution kernel module according to an embodiment of the present invention.
  • the hyperspectral Chinese medicinal material identification method of adaptive random block convolution kernel network comprises the following steps:
  • Step S100 taking hyperspectral images of Chinese herbal medicines, and constructing a hyperspectral raw data set of Chinese herbal medicines.
  • Figure 2 is a partial sample image of hyperspectral data sets of Chinese herbal medicines such as yam, Atractylodes macrocephala, Citrus aurantii, Poria cocos, and tangerine peel.
  • (a) represents a sample image of yam
  • (b) represents a sample image of The sample picture of Citrus aurantium
  • (d) shows the sample picture of Poria cocos
  • (e) shows the sample picture of tangerine peel.
  • step S100 includes:
  • Step S110 using a hyperspectral sorting instrument to acquire hyperspectral images of Chinese herbal medicines, and performing reflectance correction on the collected hyperspectral images of Chinese herbal medicines;
  • Step S120 use the corrected image as a sample of the hyperspectral data set of Chinese herbal medicines, and construct an original hyperspectral data set of Chinese herbal medicines.
  • I s is the hyperspectral image of the sth class of Chinese medicinal materials in the data set D S
  • N is the number of image pixels
  • L is the number of bands of the image
  • Y s is the medicinal material category label corresponding to the sth sample in the data set D S
  • Spectral hyperspectral sorting instrument V10E, N25E-SWIR
  • the spectral ranges are 400-1000nm and 1000-2500nm respectively.
  • Step S200 Using the optimal clustering framework to obtain the optimal band subset of the hyperspectral data set of Chinese herbal medicines, select the best characteristic bands of the hyperspectral data set of Chinese medicinal materials from the optimal band subset based on the cluster sorting strategy, and form the best band feature images.
  • the optimal band subset of the hyperspectral image of Chinese herbal medicines is obtained, and then the cluster sorting method is used to effectively select the best characteristic band from the optimal band subset, which greatly reduces the amount of data and High redundancy between bands.
  • step S200 includes:
  • Step S210 Calculating the local density and intra-cluster distance of each band of the hyperspectral data of Chinese medicinal materials, and normalizing the intra-cluster distance.
  • step S210 includes:
  • Step S211 Calculating the local density of each band of the hyperspectral data of Chinese medicinal materials, specifically:
  • D ij is the similarity matrix
  • i and j are the i and j bands of the hyperspectral data of Chinese medicinal materials respectively
  • d c is the cut-off distance of the area where each band is located.
  • Step S212 Calculate the intra-cluster distance of each band of the Chinese herbal medicine hyperspectral data, specifically:
  • D ij is the similarity matrix
  • i and j are the i and j bands of the hyperspectral data of Chinese medicinal materials respectively
  • the intra-cluster distance ⁇ max of the point k with the highest local density in the hyperspectral data of Chinese medicinal materials is:
  • Step S213 Normalize the intra-cluster distance ⁇ i , specifically:
  • ⁇ i ( ⁇ i - ⁇ min )./( ⁇ max - ⁇ min )
  • ⁇ i is the intra-cluster distance of each band
  • ⁇ min is the intra-cluster distance of the point with the smallest local density in the hyperspectral data of Chinese medicinal materials
  • ⁇ max is the intra-cluster distance of the point with the largest local density in the hyperspectral data of Chinese medicinal materials .
  • Step S220 Weighting the local density and the intra-cluster distance to calculate the contribution value of each band of the hyperspectral image of the Chinese herbal medicine.
  • step S220 is specifically:
  • R i is the contribution value of the i-th band
  • ⁇ i is the local density of the i-th band
  • ⁇ i is the intra-cluster distance of the i-th band.
  • the contribution value R (R 1 , R 2 , . . . , R L ), and R i is the contribution value of the i-th band.
  • Step S230 Divide the hyperspectral image of Chinese medicinal materials into a preset number of band subsets by the K-means++ clustering method, select the band with the largest contribution value of each band subset among the preset number of band subsets, and calculate the relationship between the band and the The similarity matrices of other band subsets are summed, and the summed value is recorded as F, and F is minimized to obtain a preset number of optimal band subsets.
  • F in step S230 is specifically:
  • w pk is the similarity matrix between the band with the largest contribution value and other band subsets.
  • Step S240 Re-select the band with the largest contribution value in each optimal band subset to obtain the best feature band to form the best band feature image.
  • Step S300 Use principal component analysis to reduce the dimensionality of the data in the original hyperspectral data set of Chinese medicinal materials, and use a random projection method to obtain random blocks from the reduced dimensionality hyperspectral data of Chinese medicinal materials as convolution kernels.
  • step S300 includes:
  • Step S310 Perform principal component analysis on the data in the original hyperspectral data set of Chinese herbal medicines for dimensionality reduction and whitening to obtain dimensionality-reduced data X p , where, N is the number of image pixels, and p is the number of principal components of the image;
  • Step S320 Use the random projection method to select M random blocks in the dimensionally reduced data as the convolution kernel P', where, P i ' is the ith random block convolution kernel, and w ⁇ w is the size of the convolution kernel.
  • the number of principal components of five hyperspectral images of Chinese medicinal materials is selected here, and the number of convolution kernels P' is set to 20, and the size is 20 ⁇ 20 pixels.
  • Step S400 Modifying the convolution kernel by using a pixel adaptive method to obtain an adaptive random block convolution kernel.
  • step S400 includes:
  • Step S410 performing bilateral filtering on the best band feature image to obtain the filtered best band feature image
  • Step S500 Using a layered network to extract features of Chinese herbal medicines by convolution with an adaptive random block convolution kernel and the best band feature image.
  • step S500 includes:
  • Step S510 Set the number of layers of the layered network as n.
  • Step S520 According to the adaptive random block convolution kernel and the optimal band feature image convolution, the features of the first layer of Chinese herbal medicines are extracted.
  • the characteristics of the first layer of Chinese herbal medicines are:
  • f 1 is the feature of the first layer of Chinese herbal medicines
  • p is the number of principal components of the hyperspectral images of medicinal materials
  • M is the number of convolution kernels.
  • Step S530 Repeat steps S300 and S400 for the characteristics of the first layer of Chinese medicinal materials to obtain the second layer of adaptive random block convolution kernel, according to the second layer of adaptive random block convolution kernel and the first layer of Chinese medicinal materials The features are extracted by convolution to obtain the features of the second layer of Chinese herbal medicines;
  • Step S540 Repeat step S530 until the features of the nth layer of Chinese medicinal materials are extracted.
  • the image data after PCA dimensionality reduction is used as the convolution kernel and an adaptive method is used to modify the convolution kernel to convolve with the subset of feature bands, so that the network has the advantage of multi-scale and effectively extracts
  • the geometric and texture features of Chinese herbal medicines also maintain the edge information of various Chinese herbal medicines.
  • Step S600 Combining the features of Chinese medicinal materials extracted by the layered network and the image data of the best band features to construct a hyperspectral training set and a test set of Chinese medicinal materials.
  • Step S700 use SVM to train the training set to obtain a classification prediction model, and predict the test set of Chinese medicinal materials based on the classification prediction model, so as to realize the identification and classification of Chinese medicinal materials.
  • the training set is trained based on SVM, and a classification prediction model is obtained, which can accurately identify various types of Chinese medicinal materials and realize non-destructive and rapid classification of Chinese medicinal materials.
  • the hyperspectral Chinese medicinal material identification method of the adaptive random block convolution kernel network firstly obtains the optimal band subset of the hyperspectral image of Chinese medicinal material based on the optimal clustering framework, and then adopts cluster sorting
  • the method effectively selects the best feature band from the optimal band subset; then uses the random projection method to use the random block extracted from the hyperspectral image of Chinese herbal medicines as the convolution kernel; then uses the pixel adaptive method to modify the convolution kernel, and based on Feature extraction is performed on the characteristic band images of Chinese herbal medicines; again, the characteristics of Chinese herbal medicines are extracted using a layered network, and combined with the hyperspectral best band image data of Chinese herbal medicines, a hyperspectral training set and test set of Chinese herbal medicines are constructed; finally, SVM is used to analyze the training set
  • the classification prediction model is obtained by training, and based on the model, the test set of Chinese herbal medicines is predicted to realize the identification and classification of Chinese herbal medicines.
  • the present invention selects the best characteristic band of hyperspectral image data of Chinese herbal medicines, and greatly reduces the amount of data while fully retaining the original information of hyperspectral images of Chinese herbal medicines;
  • the random block in the feature layer of the hyperspectral image of Chinese herbal medicine is used as the convolution kernel to fully learn the texture and geometric features of Chinese herbal medicines;
  • the pixel adaptive method is used to modify the convolution kernel, which solves the problem that the features are very sparse and difficult in high-dimensional space.
  • the fourth adopts a layered structure, combined with the characteristics of the shallow and deep layers of the hyperspectral image of Chinese herbal medicines, so that the network has the characteristics of multi-scale, and effectively extracts the characteristic information of Chinese herbal medicines.
  • the identification accuracy of Chinese medicinal materials has been greatly improved, and the identification problem of various types and complex components of Chinese medicinal materials has been solved, and it can be applied to the rapid and non-destructive identification of various Chinese medicinal materials.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Software Systems (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Health & Medical Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Probability & Statistics with Applications (AREA)
  • Medical Informatics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Image Analysis (AREA)

Abstract

本发明公开了自适应随机块卷积核网络的高光谱中药材鉴别方法，基于最优聚类框架，获得中药材高光谱图像最优波段子集，再采用集群排序方法有效地从最优波段子集中选出最佳特征波段；接着使用随机投影方法将从中药材高光谱图像中提取的随机块作为卷积核；然后使用像素自适应方法修改卷积核，并基于中药材特征波段图像进行特征提取；再次，使用分层网络提取中药材的特征，并结合中药材高光谱最佳波段影像数据，构建中药材高光谱训练集与测试集；最后使用SVM对训练集进行训练得到分类预测模型，基于该模型对中药材测试集进行预测，大幅度提高了中药材的鉴别分类精度，解决了中药材种类多样、成分复杂的鉴别难题，可适用于各类中药材的快速无损鉴别。

Description

自适应随机块卷积核网络的高光谱中药材鉴别方法
本申请要求于2021年12月24日提交中国专利局的中国专利申请的优先权，其中国专利申请的申请号为202111593705.5，发明名称为“自适应随机块卷积核网络的高光谱中药材鉴别方法”，其全部内容通过引用结合在本申请中。
技术领域
本发明属于医药高光谱智能检测分析领域，特别是涉及一种自适应随机块卷积核网络的高光谱中药材鉴别方法。
背景技术
中医药是中华民族流传千年的瑰宝，中药材作为中医药最基础的部分，其质量安全影响着中医药的疗效，甚至关系到人民的生命安全。因此，对中药材的鉴别是中医药采集、加工以及过程质量监控极为关键的一环。
传统的中药材鉴定方法有性状鉴别、理化鉴别、显微鉴别和高效液相色谱、高效液相色谱-质谱联用鉴别等方法，但这些化学分析方法检测的周期长、价格昂贵，需要大量的有机溶剂，操作复杂，且对中药材具有破坏性，不能在现场进行快速检测。近年来，近红外光谱分析技术发展迅速，已经成为世界各国药物、化合物鉴别常用手段，但中药材属于混合物体系，其组成成分多样且复杂、图谱解析难度大，无法充分鉴别各类中药材，目前仍以中药材专家的人工定性鉴别为主，因此，急需开发一种快速鉴别中药材的方法。高光谱成像技术可以同时获取被测中药材的光谱信息和空间信息，准确反映中药材的理化性质，且获取的数据信息量十分丰富，可以实现中药材的无损鉴别。
目前已有采用高光谱成像技术结合化学计量学相关算法在中药材鉴别领域的相关研究，其处理流程首先采用降噪、散射校正等方法对获取的高光谱图像进行预处理；然后采用主成分分析(PCA)、偏最小二乘判别分析(PLS-DA)等机器学习算法对高光谱数据进行建模从而鉴别中药材。但此类处理流程无法准确 建立鉴别模型，且模型也不具备普适性，高光谱数据量又十分庞大，难以提取中药材的有效特征信息，被测中药材鉴别精度不高，因此需针对种类多样、成分复杂的中药材鉴别难题，开发一种能够有效提取中药材的空-谱特征信息、可适用于各类中药材的快速无损检测方法。
发明内容
针对以上技术问题，本发明提供一种自适应随机块卷积核网络的高光谱中药材鉴别方法。
本发明解决其技术问题采用的技术方案是：
自适应随机块卷积核网络的高光谱中药材鉴别方法，方法包括以下步骤：
步骤S100：拍摄中药材高光谱图像，构建中药材高光谱原始数据集；
步骤S200：采用最优聚类框架获得中药材高光谱数据集的最优波段子集，基于集群排序策略在最优波段子集中选出中药材高光谱数据集的最佳特征波段，组成最佳波段特征影像；
步骤S300：使用主成分分析对中药材高光谱原始数据集中的数据进行降维，使用随机投影方法从降维后的中药材高光谱数据中获取随机块作为卷积核；
步骤S400：采用像素自适应方法修改卷积核，得到自适应随机块卷积核；
步骤S500：采用分层网络使用自适应随机块卷积核与最佳波段特征影像卷积提取中药材特征；
步骤S600：结合分层网络所提取的中药材特征、最佳波段特征影像数据构建中药材高光谱训练集与测试集；
步骤S700：使用SVM对训练集进行训练得到分类预测模型，基于分类预测模型对中药材高光谱测试集进行预测，实现中药材的鉴别分类。
优选地，步骤S100包括：
步骤S110：采用高光谱分选仪获取中药材的高光谱图像，并对采集的中药材高光谱图像进行反射率校正；
步骤S120：将校正后的图像作为中药材高光谱数据集的样本，构建中药材高光谱原始数据集。
优选地，步骤S200包括：
步骤S210：计算中药材高光谱数据的每个波段的局部密度和簇内距离，并对簇内距离进行归一化；
步骤S220：将局部密度与簇内距离加权计算得到中药材高光谱图像每个波段的贡献值；
步骤S230：通过K-means++聚类方法将中药材高光谱图像划分为预设数量个波段子集，选取预设数量个波段子集中每个波段子集贡献值最大的波段，分别计算该波段与其他波段子集的相似性矩阵并求和，将求和得到的值记为F，最小化F得到预设数量个最优波段子集；
步骤S240：在每个最优波段子集中重新选取贡献值最大的波段，得到最佳特征波段，组成最佳波段特征影像。
优选地，步骤S210包括：
步骤S211：计算中药材高光谱数据的每个波段的局部密度，具体为：
Figure PCTCN2022076024-appb-000001
其中，D ij为相似性矩阵，i、j分别为中药材高光谱数据第i、j个波段，d c为每个波段所在区域的截断距离；
步骤S212：计算中药材高光谱数据的每个波段的簇内距离，具体为：
Figure PCTCN2022076024-appb-000002
其中，D ij为相似性矩阵，i、j分别为中药材高光谱数据第i、j个波段，对中药材高光谱数据中局部密度最大的点k的簇内距离δ max为：
Figure PCTCN2022076024-appb-000003
步骤S213：对簇内距离δ i进行归一化，具体为：
δ i＝(δ i-δ min)./(δ max-δ min)
其中，δ i为每个波段的簇内距离，δ min为中药材高光谱数据中局部密度最小的点的簇内距离，δ max为中药材高光谱数据中局部密度最大的点的簇内距离。
优选地，步骤S220具体为：
R i＝ρ i×δ i 2
其中，R i为第i个波段的贡献值，ρ i为为第i个波段的局部密度，δ i为第i个波段的簇内距离。
优选地，步骤S240中波段子集
Figure PCTCN2022076024-appb-000004
其中，d＝(d 1,…,d k-1) T为波段子集索引向量，0<d 1<…<d k-1<L，d i为第i个波段子集的索引值。
优选地，步骤S230中F具体为：
Figure PCTCN2022076024-appb-000005
其中，w pk为贡献值最大的波段与其他波段子集的相似性矩阵。
优选地，步骤S300包括：
步骤S310：对中药材高光谱原始数据集中的数据进行主成分分析降维加白化处理得到降维后的数据X p，其中，
Figure PCTCN2022076024-appb-000006
N为影像像元数，p为影像的主成分个数；
步骤S320：使用随机投影方法在降维后的数据中选取M个随机块作为卷积核P'，其中，
Figure PCTCN2022076024-appb-000007
P i'为第i个随机块卷积核，w×w为 卷积核的大小。
优选地，步骤S400包括：
步骤S410：对最佳波段特征影像进行双边滤波得到滤波后的最佳波段特征影像；
步骤S420：用卷积核P'在滤波后的最佳波段特征影像中选取对应空间位置、大小的块P”，其中，P”＝(P 1”,P 2”,…,P p”)；
步骤S430：将块P”与卷积核P'点积得到自适应随机块卷积核P，其中，P＝(P 1,P 2,…,P p)，P i为第i个自适应随机块卷积核。
优选地，步骤S500包括：
步骤S510：设定分层网络的层数为n；
步骤S520：根据自适应随机块卷积核和最佳波段特征影像卷积提取第一层中药材的特征；
步骤S530：对第一层中药材的特征重复步骤S300和步骤S400，得到第二层的自适应随机块卷积核，根据第二层的自适应随机块卷积核和第一层中药材的特征进行卷积提取得到第二层中药材的特征；
步骤S540：重复步骤S530直至提取得到第n层中药材的特征。
上述自适应随机块卷积核网络的高光谱中药材鉴别方法，首先基于最优聚类框架，获得中药材高光谱图像最优波段子集，再采用集群排序方法有效地从最优波段子集中选出最佳特征波段；接着使用随机投影方法将从中药材高光谱图像中提取的随机块作为卷积核；然后使用像素自适应方法修改卷积核，并基于中药材特征波段图像进行特征提取；再次，使用分层网络提取中药材的特征，并结合中药材高光谱最佳波段影像数据，构建中药材高光谱训练集与测试集；最后使用SVM对训练集进行训练得到分类预测模型，基于该模型对中药材测试集进行预测，实现中药材的鉴别分类，大幅度提高了中药材的鉴别精度，解 决了中药材种类多样、成分复杂的鉴别难题，可适用于各类中药材的快速无损鉴别。
附图说明
图1为本发明一实施例提供的自适应随机块卷积核网络的高光谱中药材鉴别方法得流程图；
图2为本发明一实施例的高光谱中药材数据集部分样本示意图；
图3为本发明一实施例的自适应随机块卷积核网络模型的结构框架示意图；
图4为本发明一实施例的自适应随机块卷积核模块的结构示意图。
具体实施方式
为了使本技术领域的人员更好地理解本发明的技术方案，下面结合附图对本发明作进一步的详细说明。
在一个实施例中，如图1所示，自适应随机块卷积核网络的高光谱中药材鉴别方法，方法包括以下步骤：
步骤S100：拍摄中药材高光谱图像，构建中药材高光谱原始数据集。
具体地，准备多种不同的中药材样品，需要说明的是，该实施例中以山药、白术、枳实、茯苓、陈皮五种中药材样品进行实验，但中药材的数量和种类并不局限于此。图2即为山药、白术、枳实、茯苓、陈皮的中药材高光谱数据集部分样本图，图2中(a)表示山药的样本图，(b)表示白术的样本图，(c)表示枳实的样本图，(d)表示茯苓的样本图，(e)表示陈皮的样本图。
在一个实施例中，步骤S100包括：
步骤S110：采用高光谱分选仪获取中药材的高光谱图像，并对采集的中药材高光谱图像进行反射率校正；
步骤S120：将校正后的图像作为中药材高光谱数据集的样本，构建中药材高光谱原始数据集。
具体地，获取中药材高光谱图像，构建中药材高光谱原始数据集D S＝{(I 1,Y 1),(I 2,Y 2),…,(I S,Y S)}：采用高光谱分选仪获取中药材的高光谱图像，并对采集的中药材高光谱图像进行反射率校正，将校正后的图像作为中药材高光谱数据集的样本。其中I s为数据集D S中第s类中药材高光谱影像，
Figure PCTCN2022076024-appb-000008
N为影像像元数，L为影像的波段数，Y s为数据集D S中第s个样本对应的药材类别标签；需要说明的是，上述过程中高光谱分选仪优选采用四川双利合谱高光谱分选仪(V10E、N25E-SWIR)，光谱范围分别为400-1000nm，1000-2500nm。
步骤S200：采用最优聚类框架获得中药材高光谱数据集的最优波段子集，基于集群排序策略在最优波段子集中选出中药材高光谱数据集的最佳特征波段，组成最佳波段特征影像。
具体地，基于最优聚类框架，获得中药材高光谱图像最优波段子集，再采用集群排序方法有效地从最优波段子集中选出最佳特征波段，极大程度降低了数据量和波段间的高度冗余性。
在一个实施例中，步骤S200包括：
步骤S210：计算中药材高光谱数据的每个波段的局部密度和簇内距离，并对簇内距离进行归一化。
在一个实施例中，步骤S210包括：
步骤S211：计算中药材高光谱数据的每个波段的局部密度，具体为：
Figure PCTCN2022076024-appb-000009
其中，D ij为相似性矩阵，i、j分别为中药材高光谱数据第i、j个波段，d c为每个波段所在区域的截断距离。
步骤S212：计算中药材高光谱数据的每个波段的簇内距离，具体为：
Figure PCTCN2022076024-appb-000010
其中，D ij为相似性矩阵，i、j分别为中药材高光谱数据第i、j个波段，对中药材高光谱数据中局部密度最大的点k的簇内距离δ max为：
Figure PCTCN2022076024-appb-000011
步骤S213：对簇内距离δ i进行归一化，具体为：
δ i＝(δ i-δ min)./(δ max-δ min)
其中，δ i为每个波段的簇内距离，δ min为中药材高光谱数据中局部密度最小的点的簇内距离，δ max为中药材高光谱数据中局部密度最大的点的簇内距离。
具体地，D ij相似性矩阵为计算每个波段间的L2范数，取d c＝2％×L×(L-1)。
步骤S220：将局部密度与簇内距离加权计算得到中药材高光谱图像每个波段的贡献值。
在一个实施例中，步骤S220具体为：
R i＝ρ i×δ i 2
其中，R i为第i个波段的贡献值，ρ i为为第i个波段的局部密度，δ i为第i个波段的簇内距离。
具体地，贡献值R＝(R 1,R 2,…,R L)，R i为第i个波段的贡献值。
步骤S230：通过K-means++聚类方法将中药材高光谱图像划分为预设数量个波段子集，选取预设数量个波段子集中每个波段子集贡献值最大的波段，分别计算该波段与其他波段子集的相似性矩阵并求和，将求和得到的值记为F，最小化F得到预设数量个最优波段子集。
在一个实施例中，步骤S230中的F具体为：
Figure PCTCN2022076024-appb-000012
其中，w pk为贡献值最大的波段与其他波段子集的相似性矩阵。
步骤S240：在每个最优波段子集中重新选取贡献值最大的波段，得到最佳特征波段，组成最佳波段特征影像。
在一个实施例中，步骤S240中波段子集
Figure PCTCN2022076024-appb-000013
其中，d＝(d 1,…,d k-1) T为波段索引向量，0<d 1<…<d k-1<L，d i为第i个波段的索引。进一步地，d 0＝0，d k＝L。
步骤S300：使用主成分分析对中药材高光谱原始数据集中的数据进行降维，使用随机投影方法从降维后的中药材高光谱数据中获取随机块作为卷积核。
在一个实施例中，步骤S300包括：
步骤S310：对中药材高光谱原始数据集中的数据进行主成分分析降维加白化处理得到降维后的数据X p，其中，
Figure PCTCN2022076024-appb-000014
N为影像像元数，p为影像的主成分个数；
步骤S320：使用随机投影方法在降维后的数据中选取M个随机块作为卷积核P'，其中，
Figure PCTCN2022076024-appb-000015
P i'为第i个随机块卷积核，w×w为卷积核的大小。
具体地，这里选取5个中药材高光谱图像的主成分个数，设置卷积核P'的个数为20个，大小为20×20个像素。
步骤S400：采用像素自适应方法修改卷积核，得到自适应随机块卷积核。
在一个实施例中，步骤S400包括：
步骤S410：对最佳波段特征影像进行双边滤波得到滤波后的最佳波段特征影像；
步骤S420：用卷积核P'在滤波后的最佳波段特征影像中选取对应空间位置、 大小的块P”，其中，P”＝(P 1”,P 2”,…,P p”)；
步骤S430：将块P”与卷积核P'点积得到自适应随机块卷积核P，其中，P＝(P 1,P 2,…,P p)，P i为第i个自适应随机块卷积核。
步骤S500：采用分层网络使用自适应随机块卷积核与最佳波段特征影像卷积提取中药材特征。
在一个实施例中，步骤S500包括：
步骤S510：设定分层网络的层数为n。
具体地，该步骤设置网络的层数n＝3。
步骤S520：根据自适应随机块卷积核和最佳波段特征影像卷积提取第一层中药材的特征。
具体地，第一层中药材的特征具体为：
Figure PCTCN2022076024-appb-000016
其中，f 1为第一层中药材的特征，p为药材高光谱影像的主成分个数，M为卷积核个数。
步骤S530：对第一层中药材的特征重复步骤S300和步骤S400，得到第二层的自适应随机块卷积核，根据第二层的自适应随机块卷积核和第一层中药材的特征进行卷积提取得到第二层中药材的特征；
步骤S540：重复步骤S530直至提取得到第n层中药材的特征。
具体地，使用分层网络，以PCA降维后的图像数据作为卷积核并使用自适应方法修改卷积核，与特征波段子集卷积，使网络具有多尺度的优点，有效的提取了中药材的几何与纹理特征并保持了各类中药材的边缘信息。
步骤S600：结合分层网络所提取的中药材特征、最佳波段特征影像数据构建中药材高光谱训练集与测试集。
具体地，将步骤500中得到的特征f＝(f 1,f 2,…,f N)，再结合步骤S200得到的最佳波段特征影像K组成中药材高光谱特征数据集D f＝(f 1,f 2,…,f N,K)；将中药材高光谱特征数据集D f随机排序后，构建训练集与测试集，每类药材划分20％作为训练集，其余作为测试集。
步骤S700：使用SVM对训练集进行训练得到分类预测模型，基于分类预测模型对中药材测试集进行预测，实现中药材的鉴别分类。
具体地，结合中药材的浅层与深层特征，基于SVM对训练集进行训练，得到分类预测模型，可以准确鉴别各类中药材，实现对中药材的无损、快速分类。
上述自适应随机块卷积核网络的高光谱中药材鉴别方法，如图3、图4所示，首先基于最优聚类框架，获得中药材高光谱图像最优波段子集，再采用集群排序方法有效地从最优波段子集中选出最佳特征波段；接着使用随机投影方法将从中药材高光谱图像中提取的随机块作为卷积核；然后使用像素自适应方法修改卷积核，并基于中药材特征波段图像进行特征提取；再次，使用分层网络提取中药材的特征，并结合中药材高光谱最佳波段影像数据，构建中药材高光谱训练集与测试集；最后使用SVM对训练集进行训练得到分类预测模型，基于该模型对中药材测试集进行预测，实现中药材的鉴别分类。
与现有技术相比，本发明其一，选出中药材高光谱图像数据最佳特征波段，在充分保留了中药材高光谱图像原始信息的同时大幅度减少了数据量；其二，使用中药材高光谱图像特征层中的随机块作为卷积核，充分学习了中药材的纹理与几何特征；其三，采用像素自适应方法修改卷积核，解决了特征在高维空间中非常稀疏和不规则的痛点，且有很好的保边效果；其四采用分层结构，结合中药材高光谱图像浅层与深层的特征，使网络具有多尺度等特点，有效提取中药材的特征信息，大幅度提高了中药材的鉴别精度，解决了中药材种类多样、成分复杂的鉴别难题，可适用于各类中药材的快速无损鉴别。
以上对本发明所提供的一种自适应随机块卷积核网络的高光谱中药材鉴别方法进行了详细介绍。本文中应用了具体个例对本发明的原理及实施方式进行了阐述，以上实施例的说明只是用于帮助理解本发明的核心思想。应当指出，对于本技术领域的普通技术人员来说，在不脱离本发明原理的前提下，还可以对本发明进行若干改进和修饰，这些改进和修饰也落入本发明权利要求的保护范围内。

Claims (10)

  1. 自适应随机块卷积核网络的高光谱中药材鉴别方法，其特征在于，所述方法包括以下步骤：
    步骤S100：拍摄中药材高光谱图像，构建中药材高光谱原始数据集；
    步骤S200：采用最优聚类框架获得所述中药材高光谱原始数据集的最优波段子集，基于集群排序策略在所述最优波段子集中选出所述中药材高光谱原始数据集的最佳特征波段，组成最佳波段特征影像；
    步骤S300：使用主成分分析对所述中药材高光谱原始数据集中的数据进行降维，使用随机投影方法从降维后的中药材高光谱数据中获取随机块作为卷积核；
    步骤S400：采用像素自适应方法修改所述卷积核，得到自适应随机块卷积核；
    步骤S500：采用分层网络使用所述自适应随机块卷积核与所述最佳波段特征影像卷积提取中药材特征；
    步骤S600：结合分层网络所提取的中药材特征、所述最佳波段特征影像数据构建中药材高光谱训练集与测试集；
    步骤S700：使用SVM对训练集进行训练得到分类预测模型，基于所述分类预测模型对中药材高光谱测试集进行预测，实现中药材的鉴别分类。
  2. 根据权利要求1所述的方法，其特征在于，步骤S100包括：
    步骤S110：采用高光谱分选仪获取中药材的高光谱图像，并对采集的中药材高光谱图像进行反射率校正；
    步骤S120：将校正后的图像作为中药材高光谱数据集的样本，构建中药材高光谱原始数据集。
  3. 根据权利要求2所述的方法，其特征在于，步骤S200包括：
    步骤S210：计算中药材高光谱数据的每个波段的局部密度和簇内距离，并 对所述簇内距离进行归一化；
    步骤S220：将所述局部密度与所述簇内距离加权计算得到所述中药材高光谱图像每个波段的贡献值；
    步骤S230：通过K-means++聚类方法将所述中药材高光谱图像划分为预设数量个波段子集，选取所述预设数量个波段子集中每个波段子集贡献值最大的波段，分别计算该波段与其他波段子集的相似性矩阵并求和，将求和得到的值记为F，最小化F得到预设数量个最优波段子集；
    步骤S240：在每个最优波段子集中重新选取贡献值最大的波段，得到最佳特征波段，组成最佳波段特征影像。
  4. 根据权利要求3所述的方法，其特征在于，步骤S210包括：
    步骤S211：计算中药材高光谱数据的每个波段的局部密度，具体为：
    Figure PCTCN2022076024-appb-100001
    其中，D ij为相似性矩阵，i、j分别为中药材高光谱数据第i、j个波段，d c为每个波段所在区域的截断距离；
    步骤S212：计算中药材高光谱数据的每个波段的簇内距离，具体为：
    Figure PCTCN2022076024-appb-100002
    其中，D ij为相似性矩阵，i、j分别为中药材高光谱数据第i、j个波段，对中药材高光谱数据中局部密度最大的点k的簇内距离δ max为：
    Figure PCTCN2022076024-appb-100003
    步骤S213：对簇内距离δ i进行归一化，具体为：
    δ i＝(δ i-δ min)./(δ max-δ min)
    其中，δ i为每个波段的簇内距离，δ min为中药材高光谱数据中局部密度最小的点的簇内距离，δ max为中药材高光谱数据中局部密度最大的点的簇内距离。
  5. 根据权利要求4所述的方法，其特征在于，步骤S220具体为：
    2
    R i＝ρ i×δ i
    其中，R i为第i个波段的贡献值，ρ i为为第i个波段的局部密度，δ i为第i个波段的簇内距离。
  6. 根据权利要求5所述的方法，其特征在于，步骤S240中波段子集
    Figure PCTCN2022076024-appb-100004
    Figure PCTCN2022076024-appb-100005
    其中，d＝(d 1,…,d k-1) T为波段子集索引向量，0<d 1<…<d k-1<L，d i为第i个波段子集的索引值。
  7. 根据权利要求6所述的方法，其特征在于，步骤S230中F具体为：
    Figure PCTCN2022076024-appb-100006
    其中，w pk为贡献值最大的波段与其他波段子集的相似性矩阵。
  8. 根据权利要求7所述的方法，其特征在于，步骤S300包括：
    步骤S310：对中药材高光谱原始数据集中的数据进行主成分分析降维加白化处理得到降维后的数据X p，其中，
    Figure PCTCN2022076024-appb-100007
    N为影像像元数，p为影像的主成分个数；
    步骤S320：使用随机投影方法在所述降维后的数据中选取M个随机块作为卷积核P'，其中，
    Figure PCTCN2022076024-appb-100008
    P i'为第i个随机块卷积核，w×w为卷积核的大小。
  9. 根据权利要求8所述的方法，其特征在于，步骤S400包括：
    步骤S410：对所述最佳波段特征影像进行双边滤波得到滤波后的最佳波段特征影像；
    步骤S420：用所述卷积核P'在所述滤波后的最佳波段特征影像中选取对应空间位置、大小的块P”，其中，P”＝(P 1”,P 2”,…,P p”)；
    步骤S430：将块P”与所述卷积核P'点积得到自适应随机块卷积核P，其中，P＝(P 1,P 2,…,P p)，P i为第i个自适应随机块卷积核。
  10. 根据权利要求4所述的方法，其特征在于，步骤S500包括：
    步骤S510：设定分层网络的层数为n；
    步骤S520：根据所述自适应随机块卷积核和所述最佳波段特征影像卷积提取第一层中药材的特征；
    步骤S530：对所述第一层中药材的特征重复步骤S300和步骤S400，得到第二层的自适应随机块卷积核，根据所述第二层的自适应随机块卷积核和所述第一层中药材的特征进行卷积提取得到第二层中药材的特征；
    步骤S540：重复步骤S530直至提取得到第n层中药材的特征。
PCT/CN2022/076024 2021-12-24 2022-02-11 自适应随机块卷积核网络的高光谱中药材鉴别方法 WO2023115682A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111593705.5A CN113989525B (zh) 2021-12-24 2021-12-24 自适应随机块卷积核网络的高光谱中药材鉴别方法
CN202111593705.5 2021-12-24

Publications (1)

Publication Number Publication Date
WO2023115682A1 true WO2023115682A1 (zh) 2023-06-29

Family

ID=79734180

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/076024 WO2023115682A1 (zh) 2021-12-24 2022-02-11 自适应随机块卷积核网络的高光谱中药材鉴别方法

Country Status (2)

Country Link
CN (1) CN113989525B (zh)
WO (1) WO2023115682A1 (zh)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113989525B (zh) * 2021-12-24 2022-03-29 湖南大学 自适应随机块卷积核网络的高光谱中药材鉴别方法
CN115979973B (zh) * 2023-03-20 2023-06-16 湖南大学 一种基于双通道压缩注意力网络的高光谱中药材鉴别方法
CN116597227A (zh) * 2023-05-29 2023-08-15 广东省麦思科学仪器创新研究院 质谱图解析方法、装置、设备及存储介质

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2013099451A1 (ja) * 2011-12-28 2013-07-04 オオクマ電子株式会社 医療材料コスト算出システム
CN107092906A (zh) * 2017-05-01 2017-08-25 刘至键 一种基于深度学习的中药材识别装置
CN108734205A (zh) * 2018-04-28 2018-11-02 东北电力大学 一种针对不同品种小麦种子的单粒定点识别技术
CN109374573A (zh) * 2018-10-12 2019-02-22 乐山师范学院 基于近红外光谱分析的黄瓜表皮农药残留识别方法
CN113008817A (zh) * 2021-02-24 2021-06-22 浙江工业大学 一种基于高光谱成像技术快速鉴别苦杏仁真伪优劣的方法
CN113269196A (zh) * 2021-07-19 2021-08-17 湖南大学 一种图卷积神经网络高光谱医药成分分析的实现方法
CN113989525A (zh) * 2021-12-24 2022-01-28 湖南大学 自适应随机块卷积核网络的高光谱中药材鉴别方法

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105138861B (zh) * 2015-05-31 2018-10-12 青岛市食品药品检验研究院 一种大黄药材真伪及基源种属类别预测模型的构建方法
CN106469316B (zh) * 2016-09-07 2020-02-21 深圳大学 基于超像素级信息融合的高光谱图像的分类方法及系统
CN109344698B (zh) * 2018-08-17 2021-09-03 西安电子科技大学 基于可分离卷积和硬阈值函数的高光谱波段选择方法

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2013099451A1 (ja) * 2011-12-28 2013-07-04 オオクマ電子株式会社 医療材料コスト算出システム
CN107092906A (zh) * 2017-05-01 2017-08-25 刘至键 一种基于深度学习的中药材识别装置
CN108734205A (zh) * 2018-04-28 2018-11-02 东北电力大学 一种针对不同品种小麦种子的单粒定点识别技术
CN109374573A (zh) * 2018-10-12 2019-02-22 乐山师范学院 基于近红外光谱分析的黄瓜表皮农药残留识别方法
CN113008817A (zh) * 2021-02-24 2021-06-22 浙江工业大学 一种基于高光谱成像技术快速鉴别苦杏仁真伪优劣的方法
CN113269196A (zh) * 2021-07-19 2021-08-17 湖南大学 一种图卷积神经网络高光谱医药成分分析的实现方法
CN113989525A (zh) * 2021-12-24 2022-01-28 湖南大学 自适应随机块卷积核网络的高光谱中药材鉴别方法

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
YURONG CHEN; HUI ZHANG; YAONAN WANG; Q. M. JONATHAN WU; YIMIN YANG: "Projected Sliced Wasserstein Autoencoder-based Hyperspectral Images Anomaly Detection", ARXIV.ORG, CORNELL UNIVERSITY LIBRARY, 201 OLIN LIBRARY CORNELL UNIVERSITY ITHACA, NY 14853, 22 December 2021 (2021-12-22), 201 Olin Library Cornell University Ithaca, NY 14853, XP091119330 *
ZHAO JINKAI, LUO YUNYUN; YANG LIU; DU WEIFENG; GE WEIHONG: "Rapid Identification of Three Medicinal Plants of Curcuma wenyujin Based on Near Infrared Spectroscopy", ZHONGHUA ZHONGYIYAO XUEKAN - CHINESE ARCHIVES OF TRADITIONAL CHINESE MEDICINE, ZHONGYIYAO XUEKAN, CN, vol. 38, no. 9, 10 September 2020 (2020-09-10), CN , XP093073808, ISSN: 1673-7717, DOI: 10.13193/j.issn.1673-7717.2020.09.039 *

Also Published As

Publication number Publication date
CN113989525A (zh) 2022-01-28
CN113989525B (zh) 2022-03-29

Similar Documents

Publication Publication Date Title
WO2023115682A1 (zh) 自适应随机块卷积核网络的高光谱中药材鉴别方法
Wang et al. Automatic classification of breast cancer histopathological images based on deep feature fusion and enhanced routing
US11170502B2 (en) Method based on deep neural network to extract appearance and geometry features for pulmonary textures classification
Ahmad Deep image retrieval using artificial neural network interpolation and indexing based on similarity measurement
CN106295124B (zh) 多种图像检测技术综合分析基因子图相似概率量的方法
CN110659665B (zh) 一种异维特征的模型构建方法及图像识别方法、装置
CN111191660A (zh) 一种基于多通道协同胶囊网络的直肠癌病理学图像分类方法
CN111090764B (zh) 基于多任务学习和图卷积神经网络的影像分类方法及装置
CN110309811A (zh) 一种基于胶囊网络的高光谱图像分类方法
Hasan et al. Automated detection and characterization of colon cancer with deep convolutional neural networks
Keivani et al. Automated analysis of leaf shape, texture, and color features for plant classification.
CN109886161A (zh) 一种基于可能性聚类和卷积神经网络的道路交通标识识别方法
CN113139512B (zh) 基于残差和注意力的深度网络高光谱影像分类方法
Zhan et al. Semi-supervised classification of hyperspectral data based on generative adversarial networks and neighborhood majority voting
Zhang et al. Feature-transfer network and local background suppression for microaneurysm detection
CN109118487A (zh) 基于非下采样轮廓波变换和卷积神经网络的骨龄评估方法
Wang et al. Self-supervised Spectral-level Contrastive Learning for Hyperspectral Target Detection
George et al. Automatic scale severity assessment method in psoriasis skin images using local descriptors
Buvaneswari et al. Detection and Classification of Alzheimer’s disease from cognitive impairment with resting-state fMRI
CN113538435B (zh) 一种基于深度学习的胰腺癌病理图像分类方法及系统
CN107203779A (zh) 基于空谱信息保持的高光谱降维方法
CN114140437A (zh) 一种基于深度学习的眼底硬渗出物分割方法
CN117409407A (zh) 一种多层次感知注意力机制的细胞微核分类方法
CN107590806A (zh) 一种基于大脑医学成像的检测方法和系统
CN112085742A (zh) 一种基于上下文注意力的nafld超声视频诊断方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22909026

Country of ref document: EP

Kind code of ref document: A1