CN110738217A - Tibetan medicine urine feature classification method based on multi-scale convolution sparse coding - Google Patents

Tibetan medicine urine feature classification method based on multi-scale convolution sparse coding Download PDF

Info

Publication number
CN110738217A
CN110738217A CN201910973467.7A CN201910973467A CN110738217A CN 110738217 A CN110738217 A CN 110738217A CN 201910973467 A CN201910973467 A CN 201910973467A CN 110738217 A CN110738217 A CN 110738217A
Authority
CN
China
Prior art keywords
urine
image
tibetan medicine
vector
sparse
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.)
Pending
Application number
CN201910973467.7A
Other languages
Chinese (zh)
Inventor
刘勇国
张艺
华尔江
泽翁拥忠
降拥四郎
杨尚明
李巧勤
杜春慧
肖迪尹
杜兆威
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.)
University of Electronic Science and Technology of China
Chengdu University of Traditional Chinese Medicine
Original Assignee
University of Electronic Science and Technology of China
Chengdu University of Traditional Chinese Medicine
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 University of Electronic Science and Technology of China, Chengdu University of Traditional Chinese Medicine filed Critical University of Electronic Science and Technology of China
Priority to CN201910973467.7A priority Critical patent/CN110738217A/en
Publication of CN110738217A publication Critical patent/CN110738217A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/50—Extraction of image or video features by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis
    • 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
    • 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
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
    • G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]

Abstract

The invention provides Tibetan medicine urine feature classification methods based on multi-scale convolution sparse coding, which comprise the following steps of urine image acquisition, urine primary feature extraction, urine feature multi-scale convolution sparse coding and urine feature classification.

Description

Tibetan medicine urine feature classification method based on multi-scale convolution sparse coding
Technical Field
The invention belongs to the technical field of image processing, and particularly relates to Tibetan medicine urine feature classification methods based on multi-scale convolution sparse coding.
Background
The effective inheritance of the clinical experience and academic thinking of national medicine is the key problem of national medicine development and promotion . the nation has successively developed the key technical demonstration research of national medicine development, the curative effect evaluation and platform technical research of national medicine characteristic therapy, the national medicine document arrangement and suitable technical screening promotion project, the national famous old traditional Chinese medicine expert inheritance studio construction project (including part of national medicine) and the like, and has obtained a lot of scores on the inheritance of national medicine.
The dry chemical method is semi-quantitative, has low detection specificity, can generate false negative or false positive detection results, has definite limitation, has the biggest problem of low speed and incapability of standardization in manual microscopic examination, has definite subjectivity because the technical methods of various inspectors are not unified , and has the accuracy related to factors such as the diagnosis and treatment skill level of doctors, the external environment and the like, lacks objective evaluation standards and has low repeatability.
Disclosure of Invention
Aiming at the defects in the prior art, the Tibetan medicine urine feature classification method based on the multi-scale convolution sparse coding, provided by the invention, provides auxiliary support for the objectification of the Tibetan medicine urine diagnosis.
In order to achieve the above purpose, the invention adopts the technical scheme that:
the scheme provides Tibetan medicine urine feature classification methods based on multi-scale convolution sparse coding, which comprise the following steps:
s1, urine image acquisition: shooting images of floaters and suspended matters in the urine temperature stage of the Tibetan medicine by using a camera, extracting a plurality of images to serve as training samples, and establishing a training set of 2-type urine characteristics;
s2, extracting primary characteristics of urine: extracting local characteristics of urine floats and suspended matters of the Tibetan medicine in the training set by utilizing a direction gradient histogram;
s3, urine feature multi-scale convolution sparse coding: decomposing the vectors of the local features of the images of the Tibetan medicine urine floater and the suspended matter into linear combinations of basis vectors, and reconstructing the local features of the Tibetan medicine urine floater and the suspended matter;
s4, urine characteristic classification: and classifying the Tibetan medicine urine floaters and the suspended matters by utilizing a Support Vector Machine (SVM) according to the local characteristics of the reconstructed Tibetan medicine urine floaters and the reconstructed suspended matters, so that the Tibetan medicine urine characteristic classification based on the multi-scale convolution sparse coding is completed.
, the step S2 includes the following steps:
s201, calculating a gray value of a urine image according to component brightness in an original color image of the Tibetan medicine urine floating objects and suspended objects in the training set;
s202, performing exponential transformation processing on the gray value of the urine image by using a Gamma correction method;
s203, calculating to obtain the gradient of the urine image in the horizontal coordinate direction and the vertical coordinate direction according to the urine image gray value after the index transformation, and calculating the gradient direction value of each pixel position;
s204, dividing the urine image into windows of 6 x 6 pixels, and counting the gradient histogram of each window to obtain the HOG feature vector of each window;
s205, the HOG characteristic vectors of all windows in the urine image are connected in series to obtain the HOG characteristic vectors in the urine image, so that the extraction of local characteristics of the urine floaters and suspended matters in the Tibetan medicine is completed.
Further , the expression of the gradation value f (i, j) of the urine image in step S201 is as follows:
Figure BDA0002232860020000031
where x, y are the pixel points of the urine image and R, G, B are the components of the image.
Further , the expression of the exponential transformation process in step S202 is as follows:
(x,y)＝I(x,y)γ
wherein, x and y are pixel points of the urine image, I is an image gray function, and gamma is an image gray function index.
Further , the expression of the gradient of the urine image in the abscissa and ordinate directions in step S203 is as follows:
Gx(x,y)＝H(x+1,y)-H(x-1,y)
Gy(x,y)＝H(x,y+1)-H(x,y-1)
wherein G isx(x,y)、Gy(x, y) and H (x, y) are respectively the horizontal direction gradient, the vertical direction gradient and the pixel value of the pixel point (x, y) in the input urine image;
the expression of the gradient magnitude and gradient direction at each pixel position is as follows:
Figure BDA0002232860020000032
Figure BDA0002232860020000033
wherein a (x, y) is a gradient direction, tan-1(. cndot.) is an arctangent function.
, the step S3 includes the following steps:
s301, constructing a base vector of a dictionary by utilizing a clustering algorithm according to local features of the images of the Tibetan medicine urine floats and the suspended matters, and initializing the dictionary;
s302, decomposing the dictionary into linear combinations of basis vectors to obtain a target function of sparse coding containing input basis vectors;
s303, according to the target function in the step S302, sparse codes of the Tibetan medicine urine image features are used as input to calculate by using a multi-scale convolution sparse model to obtain a target function of the multi-scale convolution sparse codes;
s304, solving sparse feature mapping by utilizing an Iterative Shrinkage Threshold Algorithm (ISTA) according to the target function of the multi-scale convolution sparse coding;
s305, updating the filter bank of the preset convolution layer by using a gradient descent method according to sparse feature mapping so as to complete reconstruction of local features of the urine floaters and suspended matters in the Tibetan medicine, wherein:
the expression of the convolutional filter bank is as follows:
Figure BDA0002232860020000041
wherein D represents a signal having S different dimensionsA convolutional filter bank with K filters per scale, ds,kRepresents hs×hsConvolution kernel of hsRepresenting the convolution kernel size.
Further , the expression of the objective function in step S302 is as follows:
Figure BDA0002232860020000042
where j denotes the jth column vector, m denotes the m-dimensional vector of the input image, x(j)A j-th column vector representing the input image,
Figure BDA0002232860020000043
j-th column representing a sparse vector, S (-) represents a sparse cost function, i 1.. k, k represents k cluster centers, λ represents a transformation quantity, ΦiSparse representation coefficients in a representation dictionary.
, the expression of the objective function of the multi-scale convolution sparse coding in step S303 is as follows:
Figure BDA0002232860020000044
wherein x isiRepresenting input urine image vector, N representing N-dimensional urine image input vector, ds,kIndicates that the input is hs×hsConvolution kernel of hsWhich represents the size of the convolution kernel,
Figure BDA0002232860020000045
the representation contains an input image xiS × K feature maps, a denotes normalization parameters, a denotes a two-dimensional discrete convolution operation, F denotes a norm of a vector matrix, D denotes a convolution filter bank with S different scales, and K filters per scale, Z denotes a set of sparse feature maps, S denotes different scales, K denotes a filter, K1.
, the expression of classification P (x) of the medical collection urine floaters and suspended solids by the SVM in step S4 is as follows:
Figure BDA0002232860020000051
Figure BDA0002232860020000052
wherein the content of the first and second substances,a predefined function representing an input vector, x represents an input vector of a support vector machine,
Figure BDA0002232860020000054
representing the product of the components of the input vector, wiRepresenting a weight parameter, b representing a bias parameter, akRepresenting dual parameters, p representing an input vector having dimensions p, xkA k-th column vector representing the input vector.
The invention has the beneficial effects that:
(1) the invention discloses a method for classifying floating object and suspended object characteristics in the urine temperature stage of Tibetan medicine urine diagnosis based on multi-scale convolution sparse coding, which does not need to manually design the characteristics and can automatically learn filter banks with different scales. The method has the advantages that the objectivity of characteristic classification of the urine floating objects and suspended matters in the Tibetan medicine is realized, the classification accuracy is improved, and an auxiliary decision is provided for the Tibetan medicine urine diagnosis;
(2) in the invention, a proper dictionary is found for a training sample set, high-dimensional image data is converted into low-dimensional linear representation by using sparse coding, and samples are quantized by using a proper sparse representation, so that a learning task is simplified, and the complexity of a model is reduced.
Drawings
FIG. 1 is a flow chart of the method of the present invention.
Detailed Description
The following description of the embodiments of the present invention is provided to facilitate the understanding of the present invention by those skilled in the art, but it should be understood that the present invention is not limited to the scope of the embodiments, and it will be apparent to those of ordinary skill in the art that various changes may be made without departing from the spirit and scope of the present invention as defined and defined in the appended claims, and is intended to protect the inventive concepts conceived of by the present invention.
Examples
As shown in FIG. 1, the invention provides Tibetan medicine urine feature classification methods based on multi-scale convolution sparse coding, which are realized as follows:
s1, urine image acquisition: images of floaters and suspended matters in the urine temperature stage of the Tibetan medicine are shot by the camera, a plurality of images are extracted to be used as training samples, and a training set of 2 types of urine characteristics is established.
In this embodiment, 2480 images of the floaters and the suspended solids at the urine temperature stage are shot by the camera in an accumulated manner, and 400 images of the floaters and the suspended solids are correspondingly extracted as training samples according to different interference conditions such as illumination, shooting angle and shielding, so as to establish a training set of the 2-type urine characteristics.
S2, extracting primary characteristics of urine: the local features of the urine floats and suspended matters in the Tibetan medicine in the training set are extracted by utilizing the histogram of directional gradient, and the implementation method is as follows:
s201, calculating a gray value of a urine image according to component brightness in an original color image of the Tibetan medicine urine floating objects and suspended objects in the training set;
s202, performing exponential transformation processing on the gray value of the urine image by using a Gamma correction method;
s203, calculating to obtain the gradient of the urine image in the horizontal coordinate direction and the vertical coordinate direction according to the urine image gray value after the index transformation, and calculating the gradient direction value of each pixel position;
s204, dividing the urine image into windows of 6 x 6 pixels, and counting the gradient histogram of each window to obtain the HOG feature vector of each window;
s205, the HOG characteristic vectors of all windows in the urine image are connected in series to obtain the HOG characteristic vectors in the urine image, so that the extraction of local characteristics of the urine floaters and suspended matters in the Tibetan medicine is completed.
In this embodiment, the obtained urine image includes a background such as a cup containing urine, and local features which can express floats and suspended matters need to be extracted first. The method comprises the steps of firstly dividing urine floating objects and suspended object images into 6-by-6 pixel windows, then collecting direction histograms of gradients or edges of all pixel points in the windows, and finally combining the histograms to form the feature descriptor. The HOG characteristic extraction steps are as follows:
(1) carrying out image graying, namely averaging the three-component brightness in the original color image to obtain gray values:
Figure BDA0002232860020000071
wherein x and y are pixel points of the urine image, and R, G, B is the component of the image;
(2) the Gamma correction method is adopted to carry out exponential transformation on the gray value of the image, aiming at adjusting the contrast of the image, reducing the influence caused by local shadow and illumination change of the image and simultaneously inhibiting the interference of noise, and the Gamma correction formula is as follows:
(x,y)＝I(x,y)γ
wherein x and y are pixel points of the urine image, I is an image gray function, gamma is an image gray function index, and the value of gamma is 1/2;
(3) calculating the gradient of the urine image in the abscissa and ordinate directions, and simultaneously calculating the gradient direction value of each pixel position, so as to capture contour information and further weaken the interference of illumination;
Gx(x,y)＝H(x+1,y)-H(x-1,y)
Gy(x,y)＝H(x,y+1)-H(x,y-1)
wherein G isx(x,y)、Gy(x,y)、H(x,y) Respectively inputting the horizontal direction gradient, the vertical direction gradient and the pixel value of a pixel point (x, y) in the urine image;
the gradient amplitude and gradient direction at pixel point (x, y) are respectively:
Figure BDA0002232860020000072
Figure BDA0002232860020000081
wherein a (x, y) is a gradient direction, tan-1(. cndot.) is an arctangent function.
(4) Dividing the image into 6 × 6 pixel windows, counting gradient histograms (number of different gradients) of each window to form feature vectors of each window, wherein the 6 × 6 windows form blocks, and the feature vectors of all the windows in blocks are connected in series to obtain HOG feature vectors of the blocks;
(5) the HOG feature vectors of all blocks in the image are connected in series according to the column sequence to obtain the HOG feature vector of the image, and the feature vector contains local features representing floaters and suspended solids in the urine image
S3, urine feature multi-scale convolution sparse coding: decomposing the vectors of the local features of the images of the Tibetan medicine urine floater and the suspended matter into linear combinations of basis vectors, and reconstructing the local features of the Tibetan medicine urine floater and the suspended matter, wherein the implementation method comprises the following steps:
s301, constructing a base vector of a dictionary by utilizing a clustering algorithm according to local features of the images of the Tibetan medicine urine floats and the suspended matters, and initializing the dictionary;
s302, decomposing the dictionary into linear combinations of basis vectors to obtain a target function of sparse coding containing input basis vectors;
s303, according to the target function in the step S302, sparse codes of the Tibetan medicine urine image features are used as input to calculate by using a multi-scale convolution sparse model to obtain a target function of the multi-scale convolution sparse codes;
s304, solving sparse feature mapping by utilizing an Iterative Shrinkage Threshold Algorithm (ISTA) according to the target function of the multi-scale convolution sparse coding;
s305, updating the filter bank of the preset convolution layer by using a gradient descent method according to sparse feature mapping so as to complete reconstruction of local features of the urine floaters and suspended matters in the Tibetan medicine, wherein:
the expression of the convolutional filter bank is as follows:
Figure BDA0002232860020000082
where D represents a convolutional filter bank having S different scales with K filters per scale, Ds,kRepresents hs×hsConvolution kernel of hsRepresenting the convolution kernel size.
Firstly, clustering centers, also called feature vectors, are obtained through clustering learning on HOG feature codes of all training sample images to form a dictionary of the sample set, and during the sparse coding process, groups of 'over-complete' basis vectors are found from the dictionary to more efficiently represent urine features, local feature vectors of Tibetan medicine urine floats and suspended matter images are decomposed into linear combinations of the group of basis vectors, and the urine floats and the suspended matter local features are reconstructed.
Let X be ═ X1,x2,x3,...xu]T∈RU×MFor the vector set of urine images on an M-dimensional space, a clustering algorithm K-means is used for clustering urine characteristics to initialize a dictionary, the invention adopts the clustering algorithm K-means algorithm, and an objective function is as follows:
Figure BDA0002232860020000091
wherein x isuRepresenting the urine image input vector, T representing the transpose of the vector, RU×MRepresenting the dimension of the image matrix, ckDenotes the k-th cluster center, U denotes all vector columns of the input image, U is the tableShowing a certain u-vector of all vector column numbers of the input image. C ═ C1,c2,c3,...,ck]TIs a set of K cluster centers of the dictionary obtained by the clustering algorithm K-means, L representing the matrix2Norm, L2Representing the 2 norm of the urine image matrix vector, the above equation is further transformed by to obtain:
Figure BDA0002232860020000092
wherein, Card (a)u) 1 denotes a cardinality constraint, i.e. auExcept for non-zero elements, the remaining elements are all zero, | auI denotes auL of1Norm, representing the sum of absolute values of all elements, auA sparse vector is represented by a vector of pixels,
Figure BDA0002232860020000094
notation in mathematics, meaning that L is for all u, L1The 1 norm representing the image matrix vector, s · t being symbols in the mathematics, indicates that a certain condition is satisfied.
After solving the optimization problem, auSubscript with only non-zero elements therein, which represents xuWhich class clusters belong to in the dictionary1,a2,a3,...,au]T∈RUDuring the training process, the formula can be converted into an optimal solution for A and C, wherein A represents a set of nonzero elements, C represents a set of dictionaries, and C is kept unchanged during the sparse coding process, and codes X and X represent a set of input vectors, namely sparse coding, wherein A is only variable quantity.
Card(au) The constraint of 1 may result in an inaccurate fit to X, which is further increased by to weaken auA is given asuAdd L1Norm, i.e. auMiddle ladleWith fewer non-zero elements, the optimal problem eventually translates into:
Figure BDA0002232860020000101
Figure BDA0002232860020000102
where λ is a regularization parameter, ckThe center of the k-th cluster is represented,the notation in mathematics, which means that for all K, which represent the resulting K cluster centers, the sparse code X can be expressed as:
X＝arg min||y-Qx||+λ||x||
wherein y is observation sample data (i.e. input urine image), Q is dictionary, and x is sparse vector to be estimated.
Clustering all HOG characteristic vectors of all training images through the steps to obtain K clustering centers to form basis vectors of the dictionary, wherein for each images, each HOG characteristic vectors x belongs to Rn(k>n) find sets of "overcomplete" basis vectors ΦiThe sparse input vector x to be estimated is represented as a linear combination of these basis vectors:
Figure BDA0002232860020000103
wherein, aiK is the number of samples trained for the coefficients of the basis vector. For overcomplete bases, the coefficient aiTherefore, in the sparse coding algorithm, criterion "sparsity" were added to solve the degradation problem due to overcomplete.
Figure BDA0002232860020000111
Where j denotes the jth column vector, m denotes the m-dimensional vector of the input image, x(j)A j-th column vector representing the input image,
Figure BDA0002232860020000112
the j column of the sparse vector is represented, the item in the above expression is a reconstruction item, so that the sparse coding algorithm can provide linear expressions with high fitting degree for the input vector x, the second item is a 'sparse penalty' item, and S (-) is sparse cost functions for a which is far greater than zeroiMake a "penalty", the invention selects L1Normal form cost function S (a)i)＝|ai|1. Limiting | | | phi | | | non-calculation to prevent sparse penalty from becoming small2Less than a constant B, | phi | | non-woven phosphor2Represents the distance between vectors, i.e. K, K represents K cluster centers for all i | | phi | | ground2B ≦ λ is transform quantities, which are regular parameters that control the relative importance of the two equations.
Wherein x isiVector representing input urine image, ds,kRepresenting a convolution kernel with an input of hs x hs,
Figure BDA0002232860020000113
and a is a normalization parameter, which is a two-dimensional discrete convolution operation, F denotes the norm of a vector matrix, in the present invention F takes 2, 2 norms of the matrix are used, D denotes a convolutional filter bank with S different scales, K filters per scale, Z denotes a set of sparse feature maps, where K1.
The multi-scale convolution sparse model takes sparse codes of the characteristics of the Tibetan medicine urine image as input to define
Figure BDA0002232860020000114
Is a training set of N samples of dimension m X N, X representing a set of input vectors, i.e. sparse coding, XiRepresenting the components in the set of vectors X, also vectors of the input image,
Figure BDA0002232860020000115
is a convolutional filter bank with S different scales, K filters per scale, where ds,kIs hs×hsConvolution kernel of hsRepresenting the size of the convolution kernel, S representing a convolutional filter bank of different scales, K representing K filters per scale, S representing different scales, K representing a filter, and a definition
Figure BDA0002232860020000116
Is a set of sparse feature maps, ZiRepresents the ith feature map in the sparse feature map, N represents an N-dimensional urine image input vector, wherein,
Figure BDA0002232860020000117
involving sparse coding xiS x K number of feature maps of (a),
Figure BDA0002232860020000118
the representation contains an input image xiS × K feature maps, ZiRepresenting the ith feature map in the sparse feature map, the objective of multi-scale convolutional sparse coding is to sparsely code x each training image featureiDecomposition into series sparse feature mapsAnd d iss,kExtracted from the filter bank D by solving the following objective function:
wherein, satisfy
Figure BDA0002232860020000122
The terms K1, K, s1, and l represent the reconstruction error and the term l, respectively1-norm penalty, a is a normalization parameter, a is a two-dimensional discrete convolution operation, D denotes a filter bank, Z denotes a sparse feature map.
Solving the objective function through iterative optimization D and Z, fixing of the objective function, solving another of the objective function, solving the sparse feature mapping Z by adopting an Iterative Shrinkage Threshold Algorithm (ISTA), and updating the convolution kernel D by using random gradient descent.
S4, urine characteristic classification: and classifying the Tibetan medicine urine floaters and the suspended matters by utilizing a Support Vector Machine (SVM) according to the local characteristics of the reconstructed Tibetan medicine urine floaters and the reconstructed suspended matters, so that the Tibetan medicine urine characteristic classification based on the multi-scale convolution sparse coding is completed.
In this embodiment, after sparse coding of the characteristics of the medical collection urine floaters and the suspended solids in step S3, classification of the medical collection urine floaters and the suspended solids is performed by using a support vector machine SVM. The decision function p (x) is expressed as follows:
Figure BDA0002232860020000123
wherein the content of the first and second substances,is a predefined function of x, which represents the input vector of the support vector machine,
Figure BDA0002232860020000125
is the product of the x components, wiIs a weight parameter, b is a bias parameter, wiAnd b is a parameter for which the decision function is adjustable, wiThe expansion is as follows:
Figure BDA0002232860020000126
wherein, wiIs a weight parameter, akIs a dual parameter, p denotes an input vector x, x having p dimensionskThe kth column vector representing the input vector x.
The method completes the classification of the characteristics of the floating objects and the suspended matters in the Tibetan medicine urine through the steps, realizes the characteristic design without manual work, can automatically learn filter banks with different scales, realizes the objectification of the characteristic classification of the floating objects and the suspended matters in the Tibetan medicine urine, improves the classification accuracy rate, and provides an auxiliary decision for the Tibetan medicine urine diagnosis.

Claims (9)

1, Tibetan medicine urine feature classification method based on multi-scale convolution sparse coding, which is characterized by comprising the following steps:
s1, urine image acquisition: shooting images of floaters and suspended matters in the urine temperature stage of the Tibetan medicine by using a camera, extracting a plurality of images to serve as training samples, and establishing a training set of 2-type urine characteristics;
s2, extracting primary characteristics of urine: extracting local characteristics of urine floats and suspended matters of the Tibetan medicine in the training set by utilizing a direction gradient histogram;
s3, urine feature multi-scale convolution sparse coding: decomposing the vectors of the local features of the images of the Tibetan medicine urine floater and the suspended matter into linear combinations of basis vectors, and reconstructing the local features of the Tibetan medicine urine floater and the suspended matter;
s4, urine characteristic classification: and classifying the Tibetan medicine urine floaters and the suspended matters by utilizing a Support Vector Machine (SVM) according to the local characteristics of the reconstructed Tibetan medicine urine floaters and the reconstructed suspended matters, so that the Tibetan medicine urine characteristic classification based on the multi-scale convolution sparse coding is completed.
2. The method for classifying the characteristics of the Tibetan medicine urine based on the multi-scale convolution sparse coding as claimed in claim 1, wherein the step S2 comprises the following steps:
s201, calculating a gray value of a urine image according to component brightness in an original color image of the Tibetan medicine urine floating objects and suspended objects in the training set;
s202, performing exponential transformation processing on the gray value of the urine image by using a Gamma correction method;
s203, calculating to obtain the gradient of the urine image in the horizontal coordinate direction and the vertical coordinate direction according to the urine image gray value after the index transformation, and calculating the gradient direction value of each pixel position;
s204, dividing the urine image into windows of 6 x 6 pixels, and counting the gradient histogram of each window to obtain the HOG feature vector of each window;
s205, the HOG characteristic vectors of all windows in the urine image are connected in series to obtain the HOG characteristic vectors in the urine image, so that the extraction of local characteristics of the urine floaters and suspended matters in the Tibetan medicine is completed.
3. The method for classifying the characteristics of the Tibetan medicine urine based on the multi-scale convolution sparse coding as claimed in claim 2, wherein the expression of the gray value f (i, j) of the urine image in the step S201 is as follows:
Figure FDA0002232860010000021
where x, y are the pixel points of the urine image and R, G, B are the components of the image.
4. The method for classifying the characteristics of the Tibetan medicine urine based on the multi-scale convolution sparse coding as claimed in claim 2, wherein the expression of the exponential transformation process in the step S202 is as follows:
(x,y)＝I(x,y)γ
wherein, x and y are pixel points of the urine image, I is an image gray function, and gamma is an image gray function index.
5. The method for classifying the urine features of Tibetan medicine based on multi-scale convolution sparse coding according to claim 2, wherein the expression of the gradient of the urine image in the abscissa and ordinate directions in the step S203 is as follows:
Gx(x,y)＝H(x+1,y)-H(x-1,y)
Gy(x,y)＝H(x,y+1)-H(x,y-1)
wherein G isx(x,y)、Gy(x, y) and H (x, y) are respectively the horizontal direction gradient, the vertical direction gradient and the pixel value of the pixel point (x, y) in the input urine image;
the expression of the gradient magnitude and gradient direction at each pixel position is as follows:
Figure FDA0002232860010000022
Figure FDA0002232860010000023
wherein a (x, y) is a gradient direction, tan-1(. cndot.) is an arctangent function.
6. The method for classifying the characteristics of the Tibetan medicine urine based on the multi-scale convolution sparse coding as claimed in claim 1, wherein the step S3 comprises the following steps:
s301, constructing a base vector of a dictionary by utilizing a clustering algorithm according to local features of the images of the Tibetan medicine urine floats and the suspended matters, and initializing the dictionary;
s302, decomposing the dictionary into linear combinations of basis vectors to obtain a target function of sparse coding containing input basis vectors;
s303, according to the target function in the step S302, sparse codes of the Tibetan medicine urine image features are used as input to calculate by using a multi-scale convolution sparse model to obtain a target function of the multi-scale convolution sparse codes;
s304, solving sparse feature mapping by utilizing an Iterative Shrinkage Threshold Algorithm (ISTA) according to the target function of the multi-scale convolution sparse coding;
s305, updating the filter bank of the preset convolution layer by using a gradient descent method according to sparse feature mapping so as to complete reconstruction of local features of the urine floaters and suspended matters in the Tibetan medicine, wherein:
the expression of the convolutional filter bank is as follows:
where D represents a convolutional filter bank having S different scales with K filters per scale, Ds,kRepresents hs×hsConvolution kernel of hsRepresenting the convolution kernel size.
7. The method for classifying the characteristics of the Tibetan medicine urine based on the multi-scale convolution sparse coding as claimed in claim 6, wherein the expression of the objective function in the step S302 is as follows:
Figure FDA0002232860010000032
where j denotes the jth column vector, m denotes the m-dimensional vector of the input image, x(j)A j-th column vector representing the input image,j-th column representing a sparse vector, S (-) represents a sparse cost function, i 1.. k, k represents k cluster centers, λ represents a transformation quantity, ΦiSparse representation coefficients in a representation dictionary.
8. The method for classifying the characteristics of the Tibetan medicine urine based on the multi-scale convolution sparse coding as claimed in claim 6, wherein the expression of the objective function of the multi-scale convolution sparse coding in the step S303 is as follows:
Figure FDA0002232860010000034
wherein x isiRepresenting input urine image vector, N representing N-dimensional urine image input vector, ds,kIndicates that the input is hs×hsConvolution kernel of hsWhich represents the size of the convolution kernel,
Figure FDA0002232860010000041
the representation contains an input image xiS × K feature maps, a denotes normalization parameters, a denotes two-dimensional discrete convolution operation, F denotes norm of vector matrix, D denotes convolution filter bank with S different scales, and each scale has K filters, Z denotes set of sparse feature maps, S denotes different scales, K denotes filter, K denotes set of sparse feature maps, and1, K, S1, S, K denotes K filters per scale and S denotes a convolutional filter bank of a different scale.
9. The method for classifying the characteristics of the Tibetan medicine urine based on the multi-scale convolution sparse coding as claimed in claim 1, wherein the expression of the classification P (x) of the floating and suspended Tibetan medicine urine by the support vector machine SVM in the step S4 is as follows:
Figure FDA0002232860010000042
Figure FDA0002232860010000043
wherein the content of the first and second substances,
Figure FDA0002232860010000044
a predefined function representing an input vector, x represents an input vector of a support vector machine,
Figure FDA0002232860010000045
representing the product of the components of the input vector, wiRepresenting a weight parameter, b representing a bias parameter, akRepresenting dual parameters, p representing an input vector having dimensions p, xkA k-th column vector representing the input vector.
CN201910973467.7A 2019-10-14 2019-10-14 Tibetan medicine urine feature classification method based on multi-scale convolution sparse coding Pending CN110738217A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910973467.7A CN110738217A (en) 2019-10-14 2019-10-14 Tibetan medicine urine feature classification method based on multi-scale convolution sparse coding

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910973467.7A CN110738217A (en) 2019-10-14 2019-10-14 Tibetan medicine urine feature classification method based on multi-scale convolution sparse coding

Publications (1)

Publication Number Publication Date
CN110738217A true CN110738217A (en) 2020-01-31

Family

ID=69268874

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910973467.7A Pending CN110738217A (en) 2019-10-14 2019-10-14 Tibetan medicine urine feature classification method based on multi-scale convolution sparse coding

Country Status (1)

Country Link
CN (1) CN110738217A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112108263A (en) * 2020-09-09 2020-12-22 海略(连云港)科技有限公司 Fluid suspended matter classification system and method based on vortex filter and 3D U-Net network

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103018247A (en) * 2012-11-29 2013-04-03 天津理工大学 Urine cell image classification and identification system and method
CN105160314A (en) * 2015-08-29 2015-12-16 电子科技大学 Car logo classification method based on sparse coding
CN105512681A (en) * 2015-12-07 2016-04-20 北京信息科技大学 Method and system for acquiring target category picture
CN106355615A (en) * 2016-08-31 2017-01-25 惠州学院 Potato positioning method based on HOG (histogram of oriented gradient) and SVM (support vector machine)
CN109668841A (en) * 2019-01-25 2019-04-23 电子科技大学 A kind of device and method for Tibetan medicine's urine feature digital collection
CN109670396A (en) * 2018-11-06 2019-04-23 华南理工大学 A kind of interior Falls Among Old People detection method

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103018247A (en) * 2012-11-29 2013-04-03 天津理工大学 Urine cell image classification and identification system and method
CN105160314A (en) * 2015-08-29 2015-12-16 电子科技大学 Car logo classification method based on sparse coding
CN105512681A (en) * 2015-12-07 2016-04-20 北京信息科技大学 Method and system for acquiring target category picture
CN106355615A (en) * 2016-08-31 2017-01-25 惠州学院 Potato positioning method based on HOG (histogram of oriented gradient) and SVM (support vector machine)
CN109670396A (en) * 2018-11-06 2019-04-23 华南理工大学 A kind of interior Falls Among Old People detection method
CN109668841A (en) * 2019-01-25 2019-04-23 电子科技大学 A kind of device and method for Tibetan medicine's urine feature digital collection

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
HANG CHANG ET AL.: "Unsupervised Transfer Learning via Multi-Scale Convolutional Sparse Coding for Biomedical Applications", 《IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE》 *
TAIHAO LI ET AL.: "The image-based analysis and classification of urine sediments using a LeNet-5 neural network", 《COMPUTER METHODS IN BIOMECHANICS AND BIOMEDICAL ENGINEERING: IMAGING & VISUALIZATION》 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112108263A (en) * 2020-09-09 2020-12-22 海略(连云港)科技有限公司 Fluid suspended matter classification system and method based on vortex filter and 3D U-Net network

Similar Documents

Publication Publication Date Title
CN108564085B (en) Method for automatically reading of pointer type instrument
CN107194937B (en) Traditional Chinese medicine tongue picture image segmentation method in open environment
CN108537102B (en) High-resolution SAR image classification method based on sparse features and conditional random field
CN110348399B (en) Hyperspectral intelligent classification method based on prototype learning mechanism and multidimensional residual error network
CN107563433B (en) Infrared small target detection method based on convolutional neural network
CN108564025A (en) A kind of infrared image object identification method based on deformable convolutional neural networks
CN111242906B (en) Support vector data description breast image anomaly detection method
CN105069478B (en) High-spectrum remote-sensing terrain classification method based on super-pixel tensor sparse coding
CN107229917A (en) A kind of several remote sensing image general character well-marked target detection methods clustered based on iteration
CN105913081B (en) SAR image classification method based on improved PCAnet
CN111401156B (en) Image identification method based on Gabor convolution neural network
CN110490894B (en) Video foreground and background separation method based on improved low-rank sparse decomposition
CN108734199A (en) High spectrum image robust classification method based on segmentation depth characteristic and low-rank representation
CN110853070A (en) Underwater sea cucumber image segmentation method based on significance and Grabcut
CN113838064B (en) Cloud removal method based on branch GAN using multi-temporal remote sensing data
CN103246894A (en) Ground nephogram identifying method solving problem of insensitiveness in illumination
CN109190511B (en) Hyperspectral classification method based on local and structural constraint low-rank representation
CN108460400A (en) A kind of hyperspectral image classification method of combination various features information
CN108319969B (en) Brain glioma survival period prediction method and system based on sparse representation framework
CN110647977B (en) Method for optimizing Tiny-YOLO network for detecting ship target on satellite
CN107578063B (en) Image Spectral Clustering based on fast selecting landmark point
CN111047559A (en) Method for rapidly detecting abnormal area of digital pathological section
CN112784747B (en) Multi-scale eigen decomposition method for hyperspectral remote sensing image
CN110738217A (en) Tibetan medicine urine feature classification method based on multi-scale convolution sparse coding
CN116312860B (en) Agricultural product soluble solid matter prediction method based on supervised transfer learning

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
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20200131