CN105894010B - A kind of Meibomian gland functional examination method based on rough set and improved FCM algorithm - Google Patents
A kind of Meibomian gland functional examination method based on rough set and improved FCM algorithm Download PDFInfo
- Publication number
- CN105894010B CN105894010B CN201610111942.6A CN201610111942A CN105894010B CN 105894010 B CN105894010 B CN 105894010B CN 201610111942 A CN201610111942 A CN 201610111942A CN 105894010 B CN105894010 B CN 105894010B
- Authority
- CN
- China
- Prior art keywords
- meibomian gland
- attribute
- image
- attribute data
- cluster centre
- 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
Links
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
- G06F18/23211—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with adaptive number of clusters
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/211—Selection of the most significant subset of features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/24765—Rule-based classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/18—Eye characteristics, e.g. of the iris
- G06V40/193—Preprocessing; Feature extraction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/18—Eye characteristics, e.g. of the iris
- G06V40/197—Matching; Classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/03—Recognition of patterns in medical or anatomical images
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Ophthalmology & Optometry (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
- Measuring And Recording Apparatus For Diagnosis (AREA)
Abstract
The present invention relates to Intelligent Information Processing, pattern-recognition and area of medical diagnostics, specially a kind of Meibomian gland functional examination method based on rough set and improved FCM algorithm.The present invention includes three phases: wherein the first stage carries out image preprocessing, including image enhancement and image segmentation;Second stage carries out Tamura texture feature extraction;Phase III carries out classifying rules extraction, obtains typical diagnostic rule table as Meibomian gland functional examination foundation.The present invention utilizes the advantage of Rough Set Knowledge Reduction, carries out attribute reduction for the characteristic attribute of Meibomian gland form, obtains typical Meibomian gland functional diagnosis rule list.Meanwhile the information loss in order to avoid may cause in characteristic attribute progress discretization process before attribute reduction, characteristic attribute is clustered by improved FCM algorithm, farthest the information integrity of keeping characteristics attribute.The present invention has the characteristics that strong enforceability, good classification effect, recognition accuracy are high, has actual directive significance to Meibomian gland functional examination.
Description
Technical field
The present invention relates to Intelligent Information Processing, pattern-recognition and area of medical diagnostics, it is specially a kind of based on rough set and
The Meibomian gland functional examination method of improved FCM algorithm.
Background technique
Meibomian gland dysfunction (Meibomain Gland Dysfunction, MGD) is a variety of Meibomian gland dysfunctions
General name is a kind of common ophthalmology disease.The main reason for causing meibomian gland dysfunction is gland mouth obstruction, and obstruction can be straight
Connect or cause indirectly the change of Meibomian gland form.According to tarsus anadenia degree, its form can be divided into normal type, shortened
4 type, deletion form and serious loss type grades, are shown in attached drawing 1a to Fig. 1 d.Currently, meibomian gland dysfunction is often through doctor people
It is diagnosed for observation.
Textural characteristics are of great significance to image steganalysis, are ground based on psychology of the mankind to texture visual perception
Study carefully, Tamura et al. proposes 6 attributes of expression textural characteristics, is roughness respectively, contrast, direction degree, the linearity, regular
Degree and rough degree, the present invention study Meibomian gland image using these characteristic attributes.
Rough set is that Polish mathematician Pawlak proposes that attribute reduction is one of its core content, and main thought is
Redundant attributes under the premise of keeping classification and decision-making capability constant in reduction database simplify knowledge expression, improve system
The efficiency of decision and processing.In recent years, rough set is increasingly taken seriously, and in machine learning, knowledge acquisition, Analysis of Policy Making etc.
Many fields are widely used.The attribute reduction of rough set can quickly and accurately extract the effective of Meibomian gland image
Textural characteristics attribute obtains typical Meibomian gland functional diagnosis rule list.
Fuzzy C-means clustering (Fuzzy C-Means, FCM) algorithm is a kind of clustering algorithm based on division, it is a kind of
Unsupervised fuzzy recognition method, algorithm need to only provide clusters number, constantly modify repeatedly the classification of sample, cluster centre and
Sample is under the jurisdiction of degree of membership of all categories, the final objective function for obtaining optimal classification.Improved FCM algorithm proposed by the present invention,
It is possible to prevente effectively from the information loss that may cause in continuous type attribute data discretization process before rough set attribute reduction.
Summary of the invention
The present invention is proposed and a kind of based on rough set and is changed to overcome the limitation of subjective diagnosis meibomian gland dysfunction
Into the Meibomian gland functional examination method of FCM algorithm.
To achieve the above object, the present invention is achieved by the following technical solutions:
The Meibomian gland functional examination method based on rough set and improved FCM algorithm that the present invention provides a kind of, it is special it
Be in, the specific steps are as follows:
S1] image preprocessing: Meibomian gland picture quality is improved by image processing method, Meibomian gland image is made to meet naked eyes
Observation judgement or computer analysis processing, described image processing method includes image enhancement, image segmentation;
S2] Tamura texture feature extraction: choose feature of the expression quantity of Tamura textural characteristics as Meibomian gland image
Attribute, according to characteristic attribute, to treated in step S1, Meibomian gland image is studied, and the characteristic attribute is roughness, right
Than degree, direction degree, the linearity, regularity and rough degree;
S3] classifying rules extraction: using the characteristic attribute in step S2 as Meibomian gland functional examination foundation, about by attribute
Letter forms the categorised decision table simplified, specific steps are as follows:
S31] 6 kinds in step S2 continuous characteristic attributes are clustered using improved FCM algorithm: it is continuous special to 6 kinds
It levies attribute and chooses continuous type sample attribute data, clusters number and continuous type sample attribute data are inputted, using improved FCM algorithm
Return to cluster centre corresponding with continuous type sample attribute data and subordinating degree function;According to maximum membership grade principle, by changing
The continuous feature space that original continuous type sample attribute data is formed is mapped to discrete type sample attribute data into FCM algorithm
The discrete features space of formation;
S32] for the discrete type sample attribute data in the discrete features space that is obtained in step S31, using rough set
Attribute reduction is theoretical, and the core for obtaining characteristic attribute is { roughness, contrast, the linearity, rule degree };
S33] by the decision table after arrangement reduction, according to the rule of precision > 0.75 and coverage > 0.05, to feature category
The core of property is chosen, and typical Meibomian gland functional diagnosis rule list is finally obtained;
S4] Meibomian gland image to be diagnosed is inputted, step S1 to step S31 is executed, later by obtained result and step
The functional diagnosis rule list generated in S33 compares, and obtains Meibomian gland functional examination result.
Further, in step S31, every kind of characteristic attribute has respective clusters number, according to minimum target function
Value principle determines the respective clusters number of 6 kinds of characteristic attributes and cluster centre.
Further, in step S31, improved FCM algorithm clusters sample attribute data, and every kind of characteristic attribute is poly-
The specific implementation procedure of class are as follows:
S311] basic parameter, cluster centre, the subordinated-degree matrix that improved FCM algorithm is used are initialized, input sample
This attribute data collection { xi, i=1,2 ... ... n, wherein n is sample number, the number of iterations initial value T=1;
S312] it is iterated the judgement of number T < LOOP, if so, S313 is thened follow the steps, if it is not, then exporting in cluster
The heart executes step S315;
S313] subordinated-degree matrix U is updated, according to maximum membership grade principle, cluster centre is updated, by sample attribute data collection
{xiIt is divided into each new cluster centre viSample number in affiliated class, and in counting all kinds of, is denoted as ni, judge new gather
Whether class center is isolated point, if so, isolated point is put into individually set, step S313 is executed again, if it is not, then holding
Row step S314;
S314] calculating target function value, and judge whether objective function is minimum, if so, output cluster centre, executes step
Rapid S315, if it is not, being iterated number judgement in then the number of iterations T+1, return step S312.
S315] discretization sample attribute data, form discrete type sample attribute data.
Further, judge whether new cluster centre is isolated point in step S313: if belonging to new cluster centre
Sample number n in classiValue be 1,2 or niFar smaller than sample attribute data sum then illustrates that such cluster centre is isolated point.
Further, in step s 32, for the discrete type sample attribute data obtained in step S31, using rough set
Attribute reduction theory carries out reduction, and the core for obtaining characteristic attribute is { roughness, contrast, the linearity, rule degree }, specific steps
Are as follows:
S321] input step S31 obtain discrete type sample attribute data, formed information system (U, A), attribute type i
=1;
S322] by the X in discrete type sample attribute dataiRemoval, obtains new information system (U, A '), judges whether to deposit
In U/A=U/A ', if it is not, then XiFor core attributes, step S323 is executed, if so, XiFor redundant attributes, reduction attribute Xi, update
Information system (U, A) executes step S323;
S323] attribute type i+1, judge whether i > 6, if it is not, S322 is thened follow the steps, if so, obtaining core attribute set.
Further, image enhancement is carried out using high frequency emphasis filtering in step S1, the high frequency emphasis filtering is common
Multiplier and offset, transfer function H are added on the basis of high-pass filterf(u, v) are as follows:
Hf(u, v)=a+bH (u, v)
Wherein, H (u, v) is high pass filter function, and a is offset, and b is multiplier, when offset a is less than 1, high frequency
Multiplier b is greater than 1, and low-frequency component is suppressed, and radio-frequency component is enhanced.
Further, image segmentation, specific steps are carried out using local entropy filtering and morphological method in step S1 are as follows:
S11] it is filtered using Meibomian gland Local Entropy of Image, the texture image of Meibomian gland image is obtained, local entropy is pair
The regional area of n × n centered on selected pixels point carries out entropy operation, local entropy H expression formula are as follows:
Wherein, f (i, j) is exactly the pixel of this regional area of n × n, pijIt is that current pixel gray scale accounts for the total gray scale in part
Probability, local entropy is bigger, and texture difference is smaller in window, carries out Threshold segmentation according to Meibomian gland image local entropy, extracts
Target area;
S12] by morphological method in step S11 Threshold segmentation generate over-segmentation or partitioning boundary it is rough or
Cavitation is handled, and the Meibomian gland image border that makes in step S11 that treated is smooth, empty filling, obtains Meibomian gland
Segmented image.
Compared with prior art, the present invention having the beneficial effect that
Rough set and improved FCM algorithm are applied to Meibomian gland image recognition by the present invention, utilize the attribute reduction of rough set
The textural characteristics attribute big to Meibomian gland functional examination Decision Making Effect is found out, reduction falls redundant attributes, extracts most effective classification
Decision rule, to realize the intelligence and high efficiency of Meibomian gland functional examination.Meanwhile the present invention is by improved FCM algorithm
Sample attribute data is clustered, the optimum division of sample attribute is obtained, effectively maintains the integrality of sample attribute, thus
Guarantee the performance of rough set attribute reduction and categorised decision.
Using the advantage of Rough Set Knowledge Reduction, attribute reduction is carried out for 6 kinds of textural characteristics attributes of Meibomian gland form,
It extracts most effective characteristic attribute is differentiated to tarsus anadenia degree, reduction is 2 kinds the smallest to pattern-recognition Decision Making Effect
Attribute obtains typical Meibomian gland functional diagnosis rule list.Meanwhile in order to avoid continuous type sample attribute number before attribute reduction
According to the information loss may cause in discretization process, introduces improved FCM algorithm and sample attribute data is clustered, most
Big degree retains sample attribute information integrity.
During the entire process of the present invention, parameter extraction, diagnostic rule are generated, diagnostic result output is all automatically, not
It needs artificially to specify, improves the reliability and objectivity of Meibomian gland functional examination.It is demonstrated experimentally that this method has enforceability
By force, good classification effect, the features such as recognition accuracy is high have actual directive significance to Meibomian gland functional examination.
Detailed description of the invention
Fig. 1 a is Meibomian gland function normal morphology figure based on Meibomian gland functional examination of the present invention.
Fig. 1 b is that tarsus body of gland based on Meibomian gland functional examination of the present invention shortens aspect graph.
Fig. 1 c is that tarsus glandular tube based on Meibomian gland functional examination of the present invention lacks aspect graph.
Fig. 1 d is tarsus glandular tube serious loss aspect graph based on Meibomian gland functional examination of the present invention.
Fig. 2 is the work block diagram that the present invention generates Meibomian gland functional diagnosis rule list.
Fig. 3 is the flow chart that improved FCM algorithm of the present invention carries out characteristic attribute cluster.
Fig. 4 is the flow chart of Algorithm for Attribute Reduction of the present invention.
Fig. 5 is the typical diagnostic rule list that the present invention generates.
Specific embodiment
To make the objectives, technical solutions, and advantages of the present invention clearer, below in conjunction with attached drawing 1a to Fig. 5 to this hair
It is bright to be described in further detail.
Embodiment 1:
Rough set and improved FCM algorithm are applied to Meibomian gland image recognition by the present invention, utilize the attribute reduction of rough set
The textural characteristics attribute big to Meibomian gland functional examination Decision Making Effect is found out, reduction falls redundant attributes, extracts most effective classification
Decision rule, to realize the intelligence and high efficiency of Meibomian gland functional examination.
Attached drawing 1a to Fig. 1 d provides Meibomian gland form grade figure based on Meibomian gland functional examination, is followed successively by Meibomian gland function
It can normal, tarsus body of gland shortening, tarsus glandular tube missing, tarsus glandular tube serious loss.
Attached drawing 2 gives the work block diagram that the present invention generates Meibomian gland functional diagnosis rule list, including three phases: wherein
First stage carries out image preprocessing, including image enhancement and image segmentation;Second stage carries out Tamura texture feature extraction;
Phase III carries out classifying rules extraction, obtains typical diagnostic rule table as Meibomian gland functional examination foundation.
Specific embodiment the following steps are included:
S1, image preprocessing:
Including image enhancement and image segmentation, the purpose of this step is to improve image matter by corresponding image processing method
Amount make it more appropriate to the observation judgement or the analysis of computer processing of human eye.
Image enhancement, the present invention enhance image using high frequency emphasis filtering technique.So-called high frequency emphasis filters
Refer to and adds multiplier and offset, transfer function H on the basis of common high-pass filterf(u, v) are as follows:
Hf(u, v)=a+bH (u, v)
Wherein, H (u, v) is high pass filter function, and the present invention uses Gaussian filter;A is offset, and b is to multiply
Number, when offset a is less than 1, high frequency multiplier b is greater than 1, and low-frequency component is suppressed, and radio-frequency component is enhanced.
Image segmentation, the present invention combine local entropy filtering and morphological method, devise a kind of point of texture filtering
Segmentation method.It is filtered first with Local Entropy of Image, obtains texture image, local entropy is to centered on selected pixels point
The regional area of n × n carries out entropy operation, local entropy H expression formula are as follows:
Wherein, f (i, j) is exactly the pixel of this regional area of n × n, pijIt is that current pixel gray scale accounts for the total gray scale in part
Probability, local entropy is bigger, and texture difference is smaller in window, therefore carries out Threshold segmentation according to image local entropy, extracts mesh
Mark region.It, can phenomena such as containing cavity since artificial defined threshold limitation is likely to occur over-segmentation or partitioning boundary is rough
To make its smooth of the edge by morphological method, cavity filling obtains the segmented image of high quality.
S2, texture feature extraction:
The present invention studies Meibomian gland image using 6 kinds of textural characteristics attributes that Tamura et al. is proposed, is respectively
Roughness, contrast, direction degree, the linearity, regularity and rough degree, are given below the calculation of 6 kinds of textural characteristics attribute expression quantity
Method expression.
1, roughness:
Roughness can be divided into following steps and be calculated:
Firstly, calculating size in image is 2k×2kActive window in pixel mean intensity, i.e.,
K=0 in formula, 1, K, 5, g (i, j) be gray value at (i, j).
Then, the mean intensity between the window that each pixel does not overlap in the horizontal and vertical directions is calculated separately
Difference.
Eh(x, y)=| Ak(x+2k-1,y)-Ak(x-2k-1,y)|
Ev(x, y)=| Ak(x,y+2k-1)-Ak(x,y-2k-1)|
Make E value reach maximum k each pixel energy to be used to that optimum size S is arrangedbest(x, y)=2k。
Finally, roughness Fcrs can be by calculating S in entire image (m × n)bestAverage value obtain, express are as follows:
2, contrast:
Contrast is determined by the degree that black and white two parts on image grayscale dynamic range and histogram produce a polarization.The two
Factor can pass through α4=μ4/σ4It defines, wherein μ4It is four squares, σ2It is variance.Contrast Fcon may be defined as:
3, direction degree:
Image general direction can be obtained by calculating the acuity of peak value in histogram, be indicated are as follows:
P indicates that the peak value in histogram, n indicate peak value number, indicate that the peak value is included for some peak value P, W in formula
All zone of dispersion, and Φ is the maximum value of each angular interval, ΦPIt is the center of wave crest.HD(Φ) describes figure
The distribution of angle as in.
4, the linearity:
Linearity FlinIt is defined as the coincidence degree of each pixel (i, j) direction co-occurrence matrix, when calculating co-occurrence matrix
Pel spacing is d.
P in formulaDdIt is the range points of n × n local direction co-occurrence matrix.
5, regularity:
Since the texture features of whole image are not regular, so using subregion subgraph, and calculate each subgraph
Variance.4 characteristics of subgraph are integrated to measure the regularity of texture.
Freg=1-r (σcrs+σcon+σdir+σlin)
R is normalization factor, σ in formulacrs、σcon、σdir、σlinRespectively represent roughness, contrast, direction degree, the linearity
Standard deviation.
6, rough degree:
Rough degree FrghIt is the superposition of roughness and contrast, indicates are as follows:
Frgh=Fcrs+Fcon
S3, classifying rules extract:
It is substantially carried out the work that hierarchical cluster attribute, attribute reduction and diagnostic rule table generate in step s3, hierarchical cluster attribute is adopted
With improved FCM algorithm, attribute reduction and diagnostic rule table generate and use rough set theory, and the specific execution that classifying rules extracts is such as
Under:
Step 1, hierarchical cluster attribute
The present invention using Tamura feature as Meibomian gland functional examination according to and establish knowledge-representation system.6 kinds of texture spies
Sign expression quantity is 6 kinds of characteristic attributes in the present embodiment, is followed successively by roughness, contrast, direction degree, the linearity, rule degree
And rough degree, and decision attribute includes normal Meibomian gland, shortening, missing and serious loss.Meibomian gland image texture characteristic and its
Dependence between form is not easy to be understood, it is also difficult to identification is directly used in, so being needed before application to 6 kinds of samples
Attribute data carries out sliding-model control.
6 kinds of continuous characteristic attribute variables are clustered using the improved FCM algorithm of the present invention, input clusters number,
The cluster centre and subordinating degree function of update can be returned using improved FCM algorithm.According to maximum membership grade principle, pass through improvement
The continuous feature space that original continuous type sample attribute data is formed is mapped to discrete type sample attribute data shape by FCM algorithm
At discrete features space.
In the present embodiment, every kind of characteristic attribute respectively corresponds to different clusters numbers, and clusters number takes in the present embodiment
Being worth range is 3~5.The value range of clusters number is determined according to Meibomian gland form number of levels is 4 grades, final basis
Target function value minimum determines the clusters number and cluster centre of 6 kinds of characteristic attributes.
For data set { xi(i=1,2, L, n), vi(i=1,2, L, c) is the center of each cluster, and n is number of samples,
C is clusters number, μikIt is the degree of membership that k-th of sample belongs to the i-th class, Dunn is by objective function is defined as:
Wherein, U is subordinated-degree matrix, and V is cluster centre matrix;||xk-vi||2It is sample xkTo cluster centre viIt is European
Distance;M is a Weighted Index, affects subordinated-degree matrix blurring degree, and FCM algorithm is exactly based on μikAnd viIteration come
Obtain cluster centre.
In the present embodiment, by sample attribute data carry out Fuzzy processing the result is that: 6 kinds of characteristic attributes it is poly-
Class number is respectively 4,5,4,4,4,5, and according to cluster centre, each sample attribute numerical value corresponds to a discrete features attribute
Value.
The present invention clusters sample attribute data by improved FCM algorithm, obtains the optimum division of sample attribute, has
Effect maintains the integrality of sample attribute, to guarantee the performance of rough set attribute reduction and categorised decision.
Fuzzy C-means clustering (Fuzzy C-Means, FCM) algorithm is a kind of clustering algorithm based on division, and algorithm passes through
Classification, cluster centre and the sample for modifying sample repeatedly are under the jurisdiction of degree of membership of all categories, the final mesh for obtaining optimal classification
Scalar functions.FCM algorithm is more sensitive to isolated point, if containing isolated point in data, will seriously affect the Clustering Effect of algorithm,
Algorithm is set to reach local optimum.
Based on algorithm above theory, show that improved FCM algorithm carries out characteristic attribute cluster process, see attached drawing 3.
In improved FCM algorithm in the present embodiment,
A, basic parameter, cluster centre, the subordinated-degree matrix that improved FCM algorithm is used are initialized, input sample
Attribute data collection { xi, (i=1,2 ... ... n) (wherein n is sample number), the number of iterations T=1;
B, it is iterated the judgement of number T < LOOP, if so, c is thened follow the steps, if it is not, cluster centre is then exported, it is discrete
Change sample attribute data;
C, subordinated-degree matrix U is updated, according to maximum membership grade principle, cluster centre is updated, by sample data { xiDivide
To each new cluster centre viSample number in affiliated class, and in counting all kinds of, is denoted as ni, judge that new cluster centre is
No is isolated point, and class is corresponding characteristic attribute here.
By sample attribute data be divided to it is all kinds of after, the sample number n of every classiIt is different, according to niValue, class can be divided into greatly
Sample set, small sample set, if the n of large sample collectioniThe far smaller than n of total sample number or small sample setiValue is 1,2, then says
Such bright cluster centre is isolated point.
If so, isolated point is put into individually set, the sample attribute data except isolated point is clustered, if
It is no, then follow the steps d.
D, calculating target function value, and judge whether objective function is minimum, if so, output cluster centre, discretization sample
This attribute data, if it is not, being iterated number judgement in then the number of iterations T+1, return step b.
Experiment shows that isolated point and normal data value are put together and counted by traditional FCM algorithm, affects in cluster
The selection of the heart, so that final goal function reaches local optimum;After improved FCM algorithm of the invention, when cluster, will isolate
Point is put into individually set, to clustering to other data except isolated point, so that the cluster centre chosen is more
It added with effect, obtains Different categories of samples number and is evenly distributed rationally, the number of iterations significantly reduces, while keeping objective function more excellent.
Step 2, attribute reduction
For the discrete type attribute data of acquisition, reduction is carried out using the attribute reduction theory of rough set, the present invention uses
Johnson algorithm obtains the core of characteristic attribute are as follows: { roughness, contrast, the linearity, rule degree } shows for decision attribute
For, this 4 kinds of characteristic attributes keep the categorised decision ability of system enough.
Algorithm for Attribute Reduction is studied below.
1, rough set basic theories:
Theorem 1 sets the non-empty domain that U is our interested object compositions.Any subsetOne in referred to as U
Concept and scope.Family on U is divided into a knowledge base about U.
It is an equivalence relation on U that theorem 2, which sets R, and U/R indicates the set that all equivalence classes of R are constituted, [x]RIndicate packet
The R equivalence class of the U of ∈ containing element x.One knowledge base is exactly a relational system K=(U, R), wherein setting U is nonempty finite set,
Referred to as domain, R are an equivalence relations on U.
If theorem 3AndThen ∩ P (intersections of all equivalence relations in P) is also an equivalence relation, is claimed
For the indiscriminate relation on P, it is denoted as ind (P), and is had:
2, attribute reduction:
Attribute reduction is one of core content of rough set theory.Attribute in sample attribute library is not of equal importance
, certain knowledge are redundancies even in.Attribute reduction is exactly to delete under conditions of keeping knowledge-based classification ability constant
Wherein uncorrelated or unessential knowledge.There are two basic conceptions in attribute reduction: reduction and core.
It is family's equivalence relation that theorem 1, which enables R, and R ∈ R, if ind (R)=ind (R- { R }), R is referred to as unnecessary in R
's;Otherwise R is referred to as necessary in R;If each R ∈ R be it is necessary in R, referred to as R is independent;Otherwise R be referred to as according to
Bad.
Theorem 2 is setIf Q is independent, and ind (Q)=ind (P), then Q is referred to as a reduction of P.Institute in P
It is necessary to the cores that the collection of relationship composition is collectively referred to as P, are denoted as core (P).
3, decision table:
Knowledge-representation system is known as information system, is usually expressed with S=(U, A), and wherein U is the nonempty finite set of object
It closes, referred to as domain;A is the nonempty finite set of attribute.
Decision table is a kind of special knowledge-representation system.Decision determines that rule is defined as follows:
rij:des(Xi)→des(Yj),
Certainty factor μ (the X of rulei,Yj)=| Yj∩Xi|/Xi, 0 < μ (Xi,Yj)<1。
As μ (Xi,YjWhen)=1, rijIt is determining;As 0 < μ (Xi,YjWhen) < 1, rijIt is uncertain.
Rough set (Rough Set, RS) theory is built upon on the basis of classification mechanism, for portray imperfection,
Probabilistic completely new mathematical tool, it can excavate tacit knowledge and rule from mass data, not due to rough set theory
Any priori knowledge is needed, therefore is rapidly progressed application in practice.Rough set has become artificial intelligence in recent years
A newer academic hot spot in field, validity are demonstrate,proved in the successful application in many scientific and engineering fields
It is real.Currently, rough set in medical image recognition using being not much, can document for reference also more lack, but with other figures
As identification technology is compared, rough set is more suitable for handling ambiguity and uncertain stronger medical image.
Given information system S=(U, A), wherein U is non-empty domain, A=CUD, CIC indicates conditional attribute collection, D
Indicate decision kind set.Conditional attribute collection C of the invention is the characteristic attribute collection of corresponding Meibomian gland image.If
Then pos (d)=UB(X) | X ∈ U/ind (d) } it is relative positive field of the decision attribute d relative to B.
If P and Q are equivalence relation clusters, if posind(P)(ind (Q))=posind(P-{R})(ind (Q)) then claims R ∈ P
It is that P can divide out.The collection of the equivalence relation for the Q that can not be divided out in all P is collectively referred to as the Q core of P, is denoted as coreQ(P).Attribute
Core is exactly that conditional attribute concentrates attribute those of mostly important for categorised decision, has lacked them, the quality of classification will decline.
Based on algorithm above theory, obtains Algorithm for Attribute Reduction process, see attached drawing 4.
1, the discrete type sample attribute data that input hierarchical cluster attribute obtains, forms information system (U, A), attribute type i=
1;
2, sample attribute data X is removedi, new information system (U, A ') is obtained, U/A=U/A' is judged whether there is, if
It is no, then XiFor core attributes, step 3 is executed, if so, XiFor redundant attributes, reduction attribute Xi, update information system (U, A), hold
Row step 3;
3, attribute type i+1 judges whether i > 6, if it is not, 2 are thened follow the steps, if so, obtaining core attribute set.
By the decision table after arrangement reduction, the rule of precision > 0.75 and coverage > 0.05 is chosen, typical case is finally obtained
Meibomian gland functional diagnosis rule list, see attached drawing 5.
S4, Meibomian gland functional diagnosis: inputting Meibomian gland image to be diagnosed, and executes step S1 to step S31, incites somebody to action later
To result and step S33 in the diagnostic rule table that generates compare, obtain Meibomian gland functional examination result.
When executing step 31, using 4 kinds of characteristic attributes after attribute reduction: roughness, contrast, the linearity, rule degree.
It is obvious to a person skilled in the art that invention is not limited to the details of the above exemplary embodiments, Er Qie
In the case where without departing substantially from spirit or essential attributes of the invention, the present invention can be realized in other specific forms.Therefore, no matter
From the point of view of which point, the present embodiments are to be considered as illustrative and not restrictive, and the scope of the present invention is by appended power
Benefit requires rather than above description limits, it is intended that all by what is fallen within the meaning and scope of the equivalent elements of the claims
Variation is included within the present invention.
In addition, it should be understood that although this specification is described in terms of embodiments, but not each embodiment includes
One independent technical solution, this description of the specification is merely for the sake of clarity, and those skilled in the art should incite somebody to action
As a whole, the technical solutions in the various embodiments may also be suitably combined for specification, and forming those skilled in the art can
With the other embodiments of understanding.
Claims (5)
1. a kind of Meibomian gland functional examination method based on rough set and improved FCM algorithm, it is characterised in that: specific steps are such as
Under:
S1] image preprocessing: Meibomian gland picture quality is improved by image processing method, Meibomian gland image is made to meet the sight of naked eyes
Judgement or the analysis of computer processing are examined, described image processing method includes image enhancement, image segmentation;
S2] Tamura texture feature extraction: characteristic attribute of the expression quantity of Tamura textural characteristics as Meibomian gland image is chosen,
The characteristic attribute is roughness, contrast, direction degree, the linearity, regularity and rough degree, according to characteristic attribute to step S1
In treated Meibomian gland image according to step S3] to step S4] studied;
S3] classifying rules extraction: using the characteristic attribute in step S2 as Meibomian gland functional examination foundation, pass through attribute reduction shape
At the categorised decision table simplified, specific steps are as follows:
S31] 6 kinds in step S2 continuous characteristic attributes are clustered using improved FCM algorithm: to 6 kinds of continuous feature categories
Property choose continuous type sample attribute data, input clusters number and continuous type sample attribute data, returned using improved FCM algorithm
Cluster centre corresponding with continuous type sample attribute data and subordinating degree function;According to maximum membership grade principle, by improving FCM
The continuous feature space that original continuous type sample attribute data is formed is mapped to discrete type sample attribute data and is formed by algorithm
Discrete features space;
In step S31, improved FCM algorithm clusters sample attribute data, the specific execution of every kind of characteristic attribute cluster
Process are as follows:
S311] basic parameter, cluster centre, the subordinated-degree matrix that improved FCM algorithm is used are initialized, input sample category
Property data set { xi, i=1,2 ... ... n, wherein n is sample number, and the number of iterations initial value T=1, the basic parameter includes mesh
Scalar functions;
S312] it is iterated the judgement of number T < LOOP, if so, thening follow the steps S313, if it is not, then exporting cluster centre, hold
Row step S315;
S313] subordinated-degree matrix U is updated, according to maximum membership grade principle, cluster centre is updated, by sample attribute data collection { xi}
It is divided into each new cluster centre viSample number in affiliated class, and in counting all kinds of, is denoted as ni, judge in new cluster
Whether the heart is isolated point, if so, isolated point is put into individually set, step S313 is executed again, if it is not, then executing step
Rapid S314;
S314] calculating target function value, and judge whether objective function is minimum, if so, output cluster centre, executes step
S315, if it is not, being iterated number judgement in then the number of iterations T+1, return step S312;
S315] discretization sample attribute data, form discrete type sample attribute data;
S32] for the discrete type sample attribute data in the discrete features space that is obtained in step S31, using rough set attribute
Reduction theory, the core for obtaining characteristic attribute is { roughness, contrast, the linearity, rule degree }, specific steps are as follows:
S321] discrete type sample attribute data that input step S31 is obtained, it is formed information system (U, A), wherein U is object
Nonempty finite set, referred to as domain;A is the nonempty finite set of attribute;Attribute type i=1;
S322] by the X in discrete type sample attribute dataiRemoval, obtains new information system (U, A '), judges whether there is U/A
=U/A ', if it is not, then XiFor core attributes, step S323 is executed, if so, XiFor redundant attributes, reduction attribute Xi, more new information
System (U, A) executes step S323;
S323] attribute type i+1, judge whether i > 6, if it is not, S322 is thened follow the steps, if so, obtaining core attribute set;
S33] knowledge-representation system is known as information system, and decision table is a kind of special knowledge-representation system, is come with S=(U, A)
Expression, by the decision table after arrangement reduction, according to the rule of precision > 0.75 and coverage > 0.05, to the core of characteristic attribute
It is chosen, finally obtains typical Meibomian gland functional diagnosis rule list;
S4] Meibomian gland image to be diagnosed is inputted, step S1 to step S31 is executed, it later will be in obtained result and step S33
The functional diagnosis rule list of generation compares, and obtains Meibomian gland functional examination result.
2. a kind of Meibomian gland functional examination method based on rough set and improved FCM algorithm as described in claim 1, feature
It is:
In step S31, every kind of characteristic attribute has respective clusters number, determines 6 according to minimum target functional value principle
The respective clusters number of kind characteristic attribute and cluster centre.
3. a kind of Meibomian gland functional examination method based on rough set and improved FCM algorithm as described in claim 1, feature
It is:
Judge whether new cluster centre is isolated point in step S313: if sample number n in class belonging to new cluster centrei's
Value is 1,2, then illustrates that such cluster centre is isolated point.
4. a kind of Meibomian gland functional examination method based on rough set and improved FCM algorithm as described in claim 1, feature
It is:
Image enhancement is carried out using high frequency emphasis filtering in step S1, the high frequency emphasis filtering is common high-pass filter basis
It is upper to add multiplier and offset, transfer function Hf(u, v) are as follows:
Hf(u, v)=a+bH (u, v)
Wherein, H (u, v) is high pass filter function, and a is offset, and b is multiplier, when offset a is less than 1, high frequency multiplier
B is greater than 1, and low-frequency component is suppressed, and radio-frequency component is enhanced.
5. a kind of Meibomian gland functional examination method based on rough set and improved FCM algorithm as described in claim 1, feature
It is:
Image segmentation, specific steps are carried out using local entropy filtering and morphological method in step S1 are as follows:
S11] it is filtered using Meibomian gland Local Entropy of Image, the texture image of Meibomian gland image is obtained, local entropy is to choosing
The regional area of n × n centered on capture vegetarian refreshments carries out entropy operation, local entropy H expression formula are as follows:
Wherein, f (i, j) is exactly the pixel of this regional area of n × n, pijIt is the probability that current pixel gray scale accounts for the total gray scale in part,
Local entropy is bigger, and texture difference is smaller in window, carries out Threshold segmentation according to Meibomian gland image local entropy, extracts target area
Domain;
S12] it is rough or empty to the over-segmentation of Threshold segmentation generation or partitioning boundary in step S11 by morphological method
Phenomenon is handled, and the Meibomian gland image border that makes in step S11 that treated is smooth, empty filling, obtains the segmentation of Meibomian gland
Image.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610111942.6A CN105894010B (en) | 2016-02-29 | 2016-02-29 | A kind of Meibomian gland functional examination method based on rough set and improved FCM algorithm |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610111942.6A CN105894010B (en) | 2016-02-29 | 2016-02-29 | A kind of Meibomian gland functional examination method based on rough set and improved FCM algorithm |
Publications (2)
Publication Number | Publication Date |
---|---|
CN105894010A CN105894010A (en) | 2016-08-24 |
CN105894010B true CN105894010B (en) | 2019-04-19 |
Family
ID=57014162
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610111942.6A Active CN105894010B (en) | 2016-02-29 | 2016-02-29 | A kind of Meibomian gland functional examination method based on rough set and improved FCM algorithm |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN105894010B (en) |
Families Citing this family (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109064468B (en) * | 2018-08-23 | 2021-07-06 | 上海市儿童医院 | Method for quantitatively analyzing eyelid meibomian gland morphology and area by applying MATLAB |
CN109410192B (en) * | 2018-10-18 | 2020-11-03 | 首都师范大学 | Fabric defect detection method and device based on multi-texture grading fusion |
CN111157962A (en) * | 2020-01-15 | 2020-05-15 | 东方红卫星移动通信有限公司 | Interference suppression self-adaptive optimization method for OTH radar image roughness |
CN111528789B (en) * | 2020-05-27 | 2022-06-10 | 复旦大学附属眼耳鼻喉科医院 | Image detection method for evaluating definition of meibomian gland |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20100092075A1 (en) * | 2005-12-07 | 2010-04-15 | Drvision Technologies Llc | Method of directed pattern enhancement for flexible recognition |
CN102426697A (en) * | 2011-10-24 | 2012-04-25 | 西安电子科技大学 | Image segmentation method based on genetic rough set C-mean clustering |
-
2016
- 2016-02-29 CN CN201610111942.6A patent/CN105894010B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20100092075A1 (en) * | 2005-12-07 | 2010-04-15 | Drvision Technologies Llc | Method of directed pattern enhancement for flexible recognition |
CN102426697A (en) * | 2011-10-24 | 2012-04-25 | 西安电子科技大学 | Image segmentation method based on genetic rough set C-mean clustering |
Non-Patent Citations (2)
Title |
---|
An image segmentation method based on the fusion of vector quantization and edge detection with applications to medical image processing;Ailing De 等;《International Journal of Machine Learning and》;20141231;全文 * |
基于内容的医学图像检索中对 Tamura 纹理特征的算法改进;辛民宣 等;《医疗卫生装备》;20101231;全文 * |
Also Published As
Publication number | Publication date |
---|---|
CN105894010A (en) | 2016-08-24 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Li et al. | Dilated-inception net: multi-scale feature aggregation for cardiac right ventricle segmentation | |
CN107256558A (en) | The cervical cell image automatic segmentation method and system of a kind of unsupervised formula | |
CN111242174A (en) | Liver cancer image feature extraction and pathological classification method and device based on imaging omics | |
CN105894010B (en) | A kind of Meibomian gland functional examination method based on rough set and improved FCM algorithm | |
CN107977671A (en) | A kind of tongue picture sorting technique based on multitask convolutional neural networks | |
CN109635846A (en) | A kind of multiclass medical image judgment method and system | |
CN108447062A (en) | A kind of dividing method of the unconventional cell of pathological section based on multiple dimensioned mixing parted pattern | |
CN110503649A (en) | One kind is based on Multi scale U-net and the modified liver segmentation method of super-pixel | |
Singh et al. | Histopathological image analysis for breast cancer detection using cubic SVM | |
CN111767952B (en) | Interpretable lung nodule benign and malignant classification method | |
CN109064470A (en) | A kind of image partition method and device based on adaptive fuzzy clustering | |
CN111415352B (en) | Cancer metastasis panoramic pathological section analysis method based on deep cascade network | |
Song et al. | Segmentation of overlapping cytoplasm in cervical smear images via adaptive shape priors extracted from contour fragments | |
CN115100474B (en) | Thyroid gland puncture image classification method based on topological feature analysis | |
Kurmi et al. | Content-based image retrieval algorithm for nuclei segmentation in histopathology images: CBIR algorithm for histopathology image segmentation | |
Zhao et al. | Cavitary nodule segmentation in computed tomography images based on self–generating neural networks and particle swarm optimisation | |
CN115100467A (en) | Pathological full-slice image classification method based on nuclear attention network | |
Jiang et al. | A multifocus image fusion scheme based on similarity measure of transformed isosceles triangles between intuitionistic fuzzy sets | |
Song et al. | Automated segmentation of overlapping cytoplasm in cervical smear images via contour fragments | |
Purnama et al. | Follicle detection on the usg images to support determination of polycystic ovary syndrome | |
Ning et al. | Multiscale context-cascaded ensemble framework (MsC 2 EF): application to breast histopathological image | |
CN111340829B (en) | Improved DME edema regional neural network segmentation model construction method | |
CN112927242B (en) | Fast optic disc positioning method based on region positioning and group intelligent search algorithm | |
CN110084796A (en) | A kind of analysis method of complex texture CT image | |
Ding et al. | Segmentation algorithm of medical exercise rehabilitation image based on HFCNN and IoT |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |