CN1295643C - Automatic identifying method for skin micro imiage symptom - Google Patents

Abstract

The present invention relates to an automatic identifying method for a symptom a skin micro image, which is characterized in that after the skin micro image is preprocessed, a symptom partition image is input into a microcomputer; the characteristic parameter of the symptom partition image is extracted to be converted into the characteristic vector which reflects the symptom class essential, the characteristic vector is identified and calculated according to the recognition algorithm on the basis of a support vector machine, which is used for identifying the attributes of the symptom, and then, the attributes of the symptom are displayed on a microcomputer display by characters. The present invention provides the foundation for a skin diagnosis system, and can be used as the realization basis of the skin diagnosis system.

Description

Skin micro-image symptom automatic identifying method

Technical field

The present invention relates to a kind of automatic distinguishing method for image, particularly a kind of skin micro-image symptom automatic identifying method.

Background technology

The automatic identifying method of skin micro-image symptom comprises two aspects: at first be the symptom split image extraction characteristic parameter from skin micro-image, carry out Classification and Identification according to its characteristic parameter then, thereby draw the attribute of skin symptom.

In the prior art, the skin symptom Feature Extraction has: the method for imitation skin expert visual diagnostic, the size of extracting skin symptom is (with diameter or cartographic represenation of area, also available needle point, syringe needle, the grain of rice, mung bean, pea, shelled peanut, longan, lichee, walnut, egg, fist, analogy in kind such as palm), shape (ellipse, semisphere, cusp configuration, pancake, polygon, arc, annular, irregular shape), color is (red, yellow, brown, black, in vain, the normal colour of skin etc.), color and luster is (as scarlet, light red, dark red etc.), the surface is (smooth or coarse, palilate or cauliflower form, the pinnacle, the center has or not omphalos), edge (clear or fuzzy, neat or present infiltration shape, bending etc.), quality is (solid or soft, capsule, fluctuation), arrange and (be the different figure that multiple infringement forms, wire is arranged, polycyclic, gregariousness, being dispersed in property, band shape or umbrella).This method is comparatively complicated, is difficult to accurately grasp analyze.

The method of image classification has a lot, statistical pattern classification method for example, and the structure method, classification tree and neural network etc., but for higher-dimension and magnanimity classification and identification, and do not having under the situation of priori, adopt these methods all to be difficult to obtain satisfied result.The artificial neural network method is more effective in classification on a small scale, but the artificial neural network learning algorithm has its intrinsic shortcoming, as network structure determine still not to have regular reliably, reach the local optimum point easily.Therefore, for small sample, multi-class training set, artificial neural network can produce more serious over-fitting (overfitting) problem, i.e. good the and problem of value of forecasting difference of fitting result.

Summary of the invention

The object of the present invention is to provide a kind of automatic identifying method of skin micro-image symptom characteristic,, can discern the attribute of symptom automatically the pretreated symptom split image of the skin micro-image that provides.

For achieving the above object, the present invention adopts following technical proposals:

A kind of skin micro-image symptom automatic identifying method, it is characterized in that to the pretreated symptom split image of microcomputer input skin micro-image, the characteristic parameter that extracts the symptom split image changes into the proper vector of reflection symptom classification essence, this proper vector is discerned the attribute of symptom by discerning computing based on the recognizer of support vector machine, demonstrate the attribute of this symptom then on microcomputer monitor with literal, concrete steps are:

1) above-mentioned extraction symptom split image characteristic parameter changes into the concrete steps of the proper vector of reflection symptom classification essence and is:

A. extract the geometric properties of symptom:

(a) extract the area of symptom: the pixel value among the figure cut apart of adding up all kinds of symptoms is 1 number, and this number is exactly the area of symptom;

(b) extract maximum, the minimum diameter of symptom: with symptom cut apart figure with direction projection from the horizontal by the θ angle, obtain a projecting direction vector; Statistics is 1 number at this vectorial interior element, is designated as N (θ); θ increases progressively 1 ° from 0 ° to 179 ° at every turn, obtains 180 N (θ); With the maximum gauge of the maximal value among these 180 N (θ) as symptom, minimum value is as the minimum diameter of symptom;

B. extract the color characteristic of symptom:

(a) color characteristic of symptom inside extracts: the difference of average color amount of selecting the average color amount of symptom inside and background skin is as the feature of color:

I. calculate symptom tone characteristics value, be designated as Hcha:Hcha=(averHzh-averHbei) * 360; In the formula: the average color tone pitch of symptom inside is designated as averHzh; The average color tone pitch of background skin is designated as averHbei;

Ii. calculate symptom saturation degree (S) eigenwert, be designated as Scha:Scha=averSzh-averSbei; In the formula: the average staturation value of symptom inside is designated as averSzh; The average staturation value of background skin is designated as averSbei;

Iii. calculate symptom intensity (I) eigenwert, be designated as Icha:Icha=averIzh-averIbei; In the formula: the average intensity value of symptom inside is designated as averIzh; The average intensity value of background skin is designated as averIbei;

Iv. calculate symptom gray scale (G) eigenwert, be designated as Gcha:Gcha=averGzh-averGbei; In the formula: the average gray value of symptom inside is designated as averGzh; The average gray value of background skin is designated as averGbei;

(b) color characteristic at symptom edge extracts: outside the feature of difference as color of the average color amount at symptom edge and the average color amount of background skin, the difference of average color amount of also selecting the average color amount at symptom edge and symptom inside is as the feature of color:

I. the difference of the average color amount of the average color amount at symptom edge and background skin

● calculate the tone difference of symptom edge and background skin, be designated as ebHcha:ebHcha=(averHedge-averHbei) * 360; In the formula: the average tone at symptom edge is designated as averHedge; The average tone of background skin is designated as averHbei;

● calculate the saturation degree difference of symptom edge and background skin, be designated as ebScha, ebScha=averSedge-averSbei; In the formula: the average staturation at symptom edge is designated as averSedge; The average staturation of background skin is designated as averSbei;

● calculate the strength difference of symptom edge and background skin, be designated as ebIcha, ebIcha=averIedge-averIbei; In the formula: the mean intensity at symptom edge is designated as averIedge: the mean intensity of background skin is designated as averIbei;

● calculate the gray scale difference value of symptom edge and background skin, be designated as ebGcha, ebGcha=averGedge-averGbei; In the formula: the average gray at symptom edge is designated as averGedge; The average gray of background skin is designated as averGbei;

Ii. the difference of the average color amount of the average color amount at symptom edge and symptom inside

● calculate the tone difference of symptom edge and symptom inside, be designated as ezHcha, ebHcha=(averHedge-averHbei) * 360, in the formula: the average tone at symptom edge is designated as averHedge; The average tone of symptom inside is designated as averHzh;

● calculate the saturation degree difference of symptom edge and symptom inside, be designated as ezScha, ezScha=averSedge-averSzh, in the formula: the average staturation at symptom edge is designated as averSedge; The average staturation of symptom inside is designated as averSzh;

● calculate the strength difference of symptom edge and symptom inside, be designated as ezIcha, ezIcha=averIedge-averIzh, in the formula: the mean intensity at symptom edge is designated as averIedge; The mean intensity of symptom inside is designated as averIzh;

● calculate the gray scale difference value of symptom edge and symptom inside, be designated as ezGcha, ezGcha=averGedge-averGzh, in the formula: the average gray at symptom edge is designated as averGedge; The average gray of symptom inside is designated as averGzh;

(c) extraction of the shape facility of symptom and conversion computing: ((x is that image is at pixel (x, the value of y) locating, (p+q) rank square m of entire image y) to f for x, y) value to extract f _PqBe calculated as follows:

$m_{\mathrm{pq}}=\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}f(x,y)x^p{y}^q,p,q=\mathrm{0,1,2},\·\·\·$

The central moment μ of image _PqBe calculated as follows:

${\mathrm{\μ}}_{\mathrm{pq}}=\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}{(x-\stackrel{\‾}{x})}^p{(y-\stackrel{\‾}{y})}^{q}f(x,y)$

Wherein $\stackrel{\‾}{x}=\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}\mathrm{xf}(x,y)/\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}f(x,y),$ $\stackrel{\‾}{y}=\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}\mathrm{yf}(x,y)/\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}f(x,y)$

Normalization central moment η _PqBe calculated as follows:

${\mathrm{\η}}_{\mathrm{pq}}={\mathrm{\μ}}_{\mathrm{pq}}/{\mathrm{\μ}}_{00}^{\mathrm{\γ}}$

γ in the formula=(p+q+2)/2, p+q=2,3,

Can derive the calculating formula of one group of invariant moments from second moment and third moment:

φ ₁＝η ₂₀+η ₀₂

${\mathrm{\φ}}_2={({\mathrm{\η}}_{20}-{\mathrm{\η}}_{02})}^2+4{\mathrm{\η}}_{11}^2$

φ ₃＝(η ₃₀-3η ₁₂) ²+(3η ₂₁-η ₀₃) ²

φ ₄＝(η ₃₀+η ₁₂) ²+(η ₂₁+η ₀₃) ²

φ ₅＝(η ₃₀-3η ₁₂)(η ₃₀+η ₁₂)×[(η ₃₀+η ₁₂) ²-3(η ₂₁+η ₀₃) ²]+

(3η ₂₁-η ₀₃)(η ₂₁+η ₀₃)×[3(η ₃₀+η ₁₂) ²-(η ₂₁+η ₀₃) ²]

φ ₆＝(η ₂0-η ₀₂)[(η ₃₀+η ₁₂) ²-(η ₂₁+η ₀₃) ²]+

4η ₁₁(η ₃₀+η ₁₂)(η ₂₁+η ₀₃)

φ ₇＝(3η ₂₁-η ₀₃)(η ₃₀+η ₁₂)×[(η ₃₀+η ₁₂) ²-3(η ₂₁+η ₀₃) ²]+

(3η ₂₁-η ₀₃)(η ₂₁+η ₀₃)×[3(η ₀₃+η ₁₂) ²-(η ₁₂+η ₀₃) ² ]

2) concrete steps of above-mentioned identification computing are:

A. select kernel function: adopt different inner product kernel functions will cause different algorithm of support vector machine, adopt 3 more class kernel functions as follows at present:

I. polynomial kernel function: k (x, y)=(x*y+1) ^dD=1,2 ... ..

Ii.RBF (Radial Basis Function) kernel function: $K(x,y)=\exp [-\frac{{\left|\right|x-y\left|\right|}^2}{2{\mathrm{\σ}}^2}]$

Iii.Sigmoid kernel function: K (x, y)=tanh[b (x*y)-c]

B. training: for authentic specimen and false sample, make its ideal be output as 1 and-1 respectively, adopt the SVM training algorithm, promptly above-mentioned classification function formula training promptly gets a sorter based on SVM this moment to convergence;

C. identification: in the sample input category device to be identified, be output as a real number value, by preset threshold, decidable revise classification originally;

A)～c) step is for two class classification problems, and for the skin micro-image classification, we are divided into four classes, therefore, can design k (k-1)/2 sorter, and each sorter all uses two class data to train;

D. repeat b for chloasma and blackhead sample), obtain sorter SVM1;

E. repeat b for chloasma and acne sample), obtain sorter SVM2;

F. repeat b for chloasma and freckle sample), obtain sorter SVM3;

G. repeat b for freckle and blackhead sample), obtain sorter SVM4;

H. repeat b for freckle spot and acne sample), obtain sorter SVM5;

I. repeat b for blackhead and acne sample), obtain sorter SVM6;

J. in classification, adopt a kind of marking strategy:, obtain 6 sorters with training process respectively and test, be i.e. c for certain sample to be classified) step, each result is 1 minute, the score that accumulative total is of all categories;

K. selecting the pairing classification of score soprano is the classification of test data;

L. the calculating of discrimination;

M. supposition has n initial sample, takes out one of them sample at every turn and does test, does training sample for all the other n-1;

N. to this n-1 training sample, repeat d)～i) go on foot;

O. do test with the another one sample again, obtain an accuracy, 0 or 100%;

P. such process repeats n time, and each sample is all once as test sample book as a result, so last accuracy value gets is n time mean value;

Q. select the method for kernel function, pass through m exactly)～p) go on foot, select the highest kernel function of discrimination.

The present invention compared with prior art, have following conspicuous outstanding substantive distinguishing features and remarkable advantage: among the present invention, the geometric properties that symptom area, maximum gauge and the minimum diameter that the symptom split image of skin micro-image is extracted represented, one group of shape facility that invariant moments is represented that color characteristic of representing with the difference of the average color amount of the average color amount of symptom inside and background skin and computing obtain can reflect symptom classification essence more exactly; The recognition methods of a kind of skin micro-image symptom based on support vector machine (SVM) of extracting among the present invention, can solve preferably since skin micro-image have very strong non-linear, dimension is higher and belong to small sample learns these problems, image identification system calculates simple, convenient, and generalization ability is strong, is fit to very much real time processing system.The present invention lays a good foundation for the skin diagnosis system, can be as the realization foundation of skin diagnosis system.

Description of drawings

The system chart that Fig. 1 uses for one embodiment of the invention.

Fig. 2 is the micrograph image pattern of chloasma in Fig. 1 example, and wherein figure (a) is cut apart figure for symptom, and figure (b) is the symptom outline map, and figure (c) is the chloasma cromogram.

Fig. 3 is the system flowchart in Fig. 1 example

Embodiment

One embodiment of the present of invention are at the chloasma in the skin micro-image, freckle, blackhead and four kinds of typical skin symptoms of acne, provide its symptom automatic identifying method.Adopt system shown in Figure 1, carry out the extraction and the identification computing of symptom characteristic by system flow block diagram shown in Figure 3:

1. feature extraction

With the chloasma is example, and the symptom that Fig. 2 illustrates chloasma is cut apart figure, symptom outline map and symptom cromogram.Concrete steps:

1) geometric properties of calculating symptom

1.1 extract the area of symptom

Pixel value is that 1 number is 10765 in the statistical graph 2 (a), and this number is exactly the area of symptom.

1.2 extract maximum, the minimum diameter of symptom

1.2.1 Fig. 2 (a) with the direction projection from the horizontal by the θ angle, obtains a projecting direction vector;

1.2.2 statistics is 1 number at this vectorial interior element, is designated as N (θ);

1.2.3 θ increases progressively 1 ° from 0 ° to 179 ° at every turn, obtains 180 N (θ),

1.2.4 with the maximum gauge of the maximal value among these 180 N (θ) as symptom, minimum value is as the minimum diameter of symptom.The maximum gauge of Fig. 2 (a) is 151, and minimum diameter is 105.

2) symptom color characteristic

2.1 the color characteristic of Fig. 2 symptom inside extracts

2.1.1 calculate symptom tone characteristics value Hcha=(averHzh-averHbei) * 360=-220.071

2.1.2 calculate symptom saturation degree (S) eigenwert Scha=averSzh-averSbei=0.248057

2.1.3 calculate symptom intensity (I) eigenwert Icha=averIzh-averIbei=-12.7408;

2.1.4 calculate symptom gray scale (G) eigenwert Gcha=averGzh-averGbei=-24.9796;

2.2 the color characteristic at Fig. 2 symptom edge extracts

2.2.1 the difference of the average color amount at symptom edge and the average color amount of background skin

2.2.1.1 calculate the tone difference ebHcha=averHedge-averHbei=-224.289 of symptom edge and background skin

2.2.1.2 calculate the saturation degree difference e bScha=averSedge-averSbei=0.125347 of symptom edge and background skin

2.2.1.3 calculate the strength difference ebIcha=averIedge-averIbei=-8.34654 of symptom edge and background skin

2.2.1.4 calculate the gray scale difference value ebGcha=averGedge-averGbei=-16.0704 of symptom edge and background skin

2.2.2 the difference of the average color amount at symptom edge and the average color amount of symptom inside

2.2.1.1 calculate the tone difference ebHcha=averHedge-averHbei=-4.21761 of symptom edge and symptom inside

2.2.1.2 calculate the saturation degree difference e zScha=averSedge-averSzh=-0.122709 of symptom edge and symptom inside

2.2.1.3 calculate the strength difference ezIcha=averIedge-averIzh=4.39426 of symptom edge and symptom inside

2.2.1.4 calculate the gray scale difference value ezGcha=averGedge-averGzh=8.90916 of symptom edge and symptom inside

3) shape facility of Fig. 2 symptom

Calculate 7 invariant moments:

φ ₁＝η ₂₀+η ₀₂＝0.73399

${\mathrm{\φ}}_2={({\mathrm{\η}}_{20}-{\mathrm{\η}}_{02})}^2+4{\mathrm{\η}}_{11}^2=2.6432$

φ ₃＝(η ₃₀-3η ₁₂) ²+(3η ₂₁-η ₀₃) ²＝3.2293

φ ₄＝(η ₃₀+η ₁₂) ²+(η ₂₁+η ₀₃) ²＝4.61191

φ ₅＝(η ₃₀-3η ₁₂)(η ₃₀+η ₁₂)×[(η ₃₀+η ₁₂) ²-3(η ₂₁+η ₀₃) ²]+

(3η ₂₁-η ₀₃)(η ₂₁+η ₀₃)×[3(η ₃₀+η ₁₂) ²-(η ₂₁+η ₀₃) ²]＝7.87397

φ ₆＝(η ₂₀-η02)[(η ₃₀+η ₁₂) ²-(η ₂₁+η ₀₃) ²]+

4η ₁₁(η ₃₀+η ₁₂)(η ₂₁+η ₀₃)＝6.10877

φ ₇＝(3η ₂₁-η ₀₃)(η ₃₀+η ₁₂)×[(η ₃₀+η ₁₂) ²-3(η ₂₁+η ₀₃) ²]+

(3η ₂₁-η ₀₃)(η ₂₁+η ₀₃)×[3(η ₀₃+η ₁₂) ²-(η ₁₂+η ₀₃) ²]＝7.78115

Like this, the chloasma image just is converted to 22 dimensional feature vectors, as [10,765 151 105-220.0710.248057-12.7408-24.9796-224.289 0.125347-8.34654-16.0704-4.21761-0.122709 4.39426 8.90916 0.73399 2.6432 3.2293 4.611917.87397 6.10877 7.78115]

2 skin micro-image recognizers based on support vector machine (SVM)

We have collected 289 skin micro-images, under relevant professional's guidance, sample are instructed classification, and random division is two groups, are used separately as training and testing sample (159 examples/130 examples), each sample calculation its 22 eigenwerts.Like this, training sample just converts 159 22 dimensional vectors to.

Make up 6 sorters, each sorter all uses two class data to train, and during training, has attempted three class kernel functions: (1) linear kernel function (linear); (2) polynomial kernel function (polynomial); (3) warp-wise base kernel function (radial basis) is selected the kernel function that obtains best LOOCV (leaving-one method) accuracy.Through experiment, this patent adopts first kind kernel function as algorithm of support vector machine, promptly k (x, y)=(x*y+1) ³, can make the sorter discrimination the highest, result such as table 1.

Several different discriminations of kernel function under different parameters of table 1

Kernel function	RBF	linea r	Poly
										Paramete r	Do not have	Do not have	D＝1	D＝2	D＝3	D＝4	D＝5	D＝6、7	D＝8、9
Leaving-one method discrimination/%	55.88	79.41	79.41	82.54	88.24	85.29	58.82	11.76	2.94

Table 2 has provided with training data (159 example) and has trained a SVM, at last test sample book (130 example) is predicted sorting result.

Table 2 svm classifier device is to the test sample book discrimination

Sample class	Chloasma (30)	Freckle (23)	Blackhead (16)	Acne (32)	Normal skin (11)
						Nicety of grading/%	88.1	90.0	86.2	85.2	100.0

Claims (1)

1. a skin microscopic image symptom automatic recognition method, it is characterized in that to microcomputer input the symptom segmentation image after skin microscopic image preprocessing, extract the characteristic parameter of symptom segmentation image and convert into the feature vector reflecting symptom classification essence, the The eigenvector performs the recognition operation based on the recognition algorithm of the support vector machine to identify the attribute of the symptom, and then displays the attribute of the symptom in text on the computer display. The specific steps are: 1) The specific steps of converting the feature parameter of the extracted symptom segmentation image into the feature vector reflecting the essence of symptom classification are: a. Extract geometric features of symptoms: (a) Extract the area of symptoms: count the number of pixels with a value of 1 in the segmentation map of various symptoms, and this number is the area of symptoms; (b) Extract the maximum and minimum diameters of the symptoms: Project the segmentation map of the symptoms in a direction that forms an angle θ with the horizontal direction to obtain a projection direction vector; count the number of elements in this vector that are 1, and record it as N(θ ); θ increases by 1° each time from 0° to 179°, and 180 N(θ) are obtained; the maximum value among these 180 N(θ) is used as the maximum diameter of the symptom, and the minimum value is used as the minimum diameter of the symptom; b. Extract color features of symptoms: (a) Color feature extraction inside the symptom: select the difference between the average color amount inside the symptom and the average color amount of the background skin as the feature of the color: i. Calculate the characteristic value of the symptom hue, which is recorded as Hcha: Hcha=(averHzh-averHbei)*360; in the formula: the average hue value inside the symptom, which is recorded as averHzh; the average hue value of the background skin, which is recorded as averHbei; ii. Calculate the symptom saturation (S) eigenvalue, denoted as Scha: Scha=averSzh-averSbei; In the formula: the average saturation value inside the symptom, denoted as averSzh; The average saturation value of the background skin, denoted as averSbei; iii. calculate symptom intensity (I) eigenvalue, denoted as Icha: Icha=averIzh-averIbei; In the formula: the average intensity value inside the symptom, denoted as averIzh; The average intensity value of background skin, denoted as averIbei; iv. Calculate the symptom gray scale (G) eigenvalue, be recorded as Gcha: Gcha=averGzh-averGbei; In the formula: the average gray value of symptom inside, be recorded as averGzh; The average gray value of background skin, be recorded as averGbei; (b) Color Feature Extraction of Symptom Edge: In addition to the difference between the average color amount of the symptom edge and the average color amount of the background skin as the feature of the color, the difference between the average color amount of the symptom edge and the average color amount inside the symptom is also selected As a feature of color: i. The difference between the average color volume of the symptom edge and the average color volume of the background skin Calculate the hue difference between symptom edge and background skin, which is recorded as ebHcha: ebHcha=(averHedge-averHbei)*360; where: the average hue of symptom edge is recorded as averHedge; the average hue of background skin is recorded as averHbei; ●Calculate the saturation difference between the symptom edge and the background skin, denoted as ebScha, ebScha=averSedge-averSbei; where: the average saturation of the symptom edge, denoted as averSedge; the average saturation of the background skin, denoted as averSbei; ●Calculate the intensity difference between the symptom edge and the background skin, which is recorded as ebIcha, ebIcha=averIedge-averIbei; where: the average intensity of the symptom edge, which is recorded as averIedge: the average intensity of the background skin, which is recorded as averIbei; ●Calculate the gray level difference between the symptom edge and the background skin, which is recorded as ebGcha, ebGcha=averGedge-averGbei; where: the average gray value of the symptom edge is recorded as averGedge; the average gray value of the background skin is recorded as averGbei; ii. The difference between the average amount of color at the edge of the symptom and the average amount of color inside the symptom ● Calculate the hue difference between the edge of the symptom and the interior of the symptom, and record it as ezHcha, ebHcha=(averHedge-averHbei)*360, where: the average hue of the edge of the symptom is recorded as averHedge; the average hue of the interior of the symptom is recorded as averHzh; ● Calculate the saturation difference between the edge of the symptom and the interior of the symptom, which is recorded as ezScha, ezScha=averSedge-averSzh, where: the average saturation of the edge of the symptom is recorded as averSedge; the average saturation of the interior of the symptom is recorded as averSzh; ● Calculate the intensity difference between the edge of the symptom and the inside of the symptom, which is recorded as ezIcha, ezIcha=averIedge-averIzh, where: the average intensity of the edge of the symptom is recorded as averIedge; the average intensity of the inside of the symptom is recorded as averIzh; ●Calculate the gray level difference between the edge of the symptom and the interior of the symptom, which is recorded as ezGcha, ezGcha=averGedge-averGzh, where: the average gray level of the symptom edge is recorded as averGedge; the average gray level of the symptom interior is recorded as averGzh; (c) Extraction and conversion operation of shape features of symptoms: extract f(x, y) value, f(x, y) is the value of the image at the pixel (x, y), and (p+q) of the entire image The order moment m _pq is calculated according to the following formula: ${\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; pq</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; x</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}\underset{\mathrm{<span\; class="notranslate">\; the\; y</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; p</span>}}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0,1,2</span>}<span\; class="notranslate">\; \&CenterDot;</span><span\; class="notranslate">\; \&CenterDot;</span><span\; class="notranslate">\; \&Center\; Dot;</span>$ The central moment μ _pq of the image is calculated as follows: ${\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; pq</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; x</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}\underset{\mathrm{<span\; class="notranslate">\; the\; y</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; p</span>}}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; -</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; q</span>}}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>$ in $\stackrel{\&OverBar;}{x}=\underset{x}{\mathrm{\&Sigma;}}\underset{\mathrm{the\; y}}{\mathrm{\&Sigma;}}\mathrm{xf}(x,\mathrm{the\; y})/\underset{x}{\mathrm{\&Sigma;}}\underset{\mathrm{the\; y}}{\mathrm{\&Sigma;}}f(x,\mathrm{the\; y}),\stackrel{\&OverBar;}{\mathrm{the\; y}}=\underset{x}{\mathrm{\&Sigma;}}\underset{\mathrm{the\; y}}{\mathrm{\&Sigma;}}\mathrm{yf}(x,\mathrm{the\; y})/\underset{x}{\mathrm{\&Sigma;}}\underset{\mathrm{the\; y}}{\mathrm{\&Sigma;}}f(x,\mathrm{the\; y})$ The normalized central moment η _pq is calculated according to the following formula: ${\mathrm{<span\; class="notranslate">\; \&eta;</span>}}_{\mathrm{<span\; class="notranslate">\; pq</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; pq</span>}}<span\; class="notranslate">\; /</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; 00</span>}}^{\mathrm{<span\; class="notranslate">\; \&gamma;</span>}}$ In the formula, γ=(p+q+2)/2, p+q=2, 3,... A set of invariant moments can be derived from the second and third moments: φ ₁ =η ₂₀ +η ₀₂ ${\mathrm{<span\; class="notranslate">\; \&phi;</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&eta;</span>}}_{\mathrm{<span\; class="notranslate">\; 20</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&eta;</span>}}_{\mathrm{<span\; class="notranslate">\; 02</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 4</span>}{\mathrm{<span\; class="notranslate">\; \&eta;</span>}}_{\mathrm{<span\; class="notranslate">\; 11</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$ φ ₃ ＝(η ₃₀ -3η ₁₂ ) ² +(3η ₂₁ -η ₀₃ ) ² φ ₄ ＝(η ₃₀ +η ₁₂ ) ² +(η ₂₁ +η ₀₃ ) ² φ ₅ =(η ₃₀ -3η ₁₂ )(η ₃₀ +η ₁₂ )×[(η ₃₀ +η ₁₂ ) ² -3(η ₂₁ +η ₀₃ ) ² ]+(3η ₂₁ -η ₀₃ )(η ₂₁ + η ₀₃ )×[3(η ₃₀ +η ₁₂ ) ² -(η ₂₁ +η ₀₃ ) ² ] φ ₆ =(η ₂₀ -η ₀₂ )[(η ₃₀ +η ₁₂ ) ² -(η ₂₁ +η ₀₃ ) ² ]+4η ₁₁ (η ₃₀ +η ₁₂ )(η ₂₁ +η ₀₃ ) φ ₇ ＝(3η ₂₁ -η ₀₃ )(η ₃₀ +η ₁₂ )×[(η ₃₀ +η ₁₂ ) ² -3(η ₂₁ +η ₀₃ ) ² ]+(3η ₂₁ -η ₀₃ )(η ₂₁ + η ₀₃ )×[3(η ₀₃ +η ₁₂ ) ² -(η ₁₂ +η ₀₃ ) ² ] 2) The specific steps of the described recognition operation are: a Select the kernel function: Using different inner product kernel functions will lead to different support vector machine algorithms. At present, more three types of kernel functions are used as follows: i. Polynomial kernel function: k(x, y)=(x*y+1) ^d d=1, 2..... ii. RBF (Radial Basis Function) kernel function: $K=(x,\mathrm{the\; y})=\mathrm{epx}[-\frac{{\left|\right|x-\mathrm{the\; y}\left|\right|}^2}{2{\mathrm{\&sigma;}}^2}]$ iii.Sigmoid kernel function: K(x, y)=tanh[b(x*y)-c] b Training: For the real sample and the fake sample, the ideal output is set to 1 and -1 respectively, and the SVM training algorithm is used, that is, the above classification function formula is trained to converge, and a classifier based on SVM is obtained at this time; c. Identification: The sample to be identified is input into the classifier, and the output is a real value, and the category of the sample can be determined through the set threshold; Steps a) to c) are for two types of classification problems, and for the classification of skin microscopic images, they are divided into four types. Therefore, k(k-1)/2 classifiers can be designed, and each classifier uses two types of data to train; d. Repeat b) for the chloasma and blackhead samples to obtain the classifier SVM1; e. Repeat b) for the chloasma and acne samples to obtain classifier SVM2; f. Repeat b) for chloasma and freckle samples to obtain classifier SVM3; g. Repeat b) for freckles and blackhead samples to obtain classifier SVM4; h. Repeat b) for freckles and acne samples to obtain classifier SVM5; i. Repeat b) for blackhead and acne samples to obtain classifier SVM6; j. When classifying, adopt a scoring strategy: for a certain sample to be classified, use the training process to obtain 6 classifiers for testing, that is, step c), each result is 1 point, and each category is accumulated Score; k. Select the category corresponding to the highest scorer as the category of the test data; l. Calculation of recognition rate; m. Assuming that there are n initial samples, one sample is taken out for testing each time, and the remaining n-1 are used as training samples; n. Repeat steps d) to i) for the n-1 training samples; o. Test with another sample to get an accuracy, 0 or 100%; p. This process is repeated n times, and as a result, each sample is used as a test sample once, so the final accuracy value is the average value of n times; q. The method of selecting the kernel function is to select the kernel function with the highest recognition rate through steps m) to p).

Images