CN103870838A - Eye fundus image characteristics extraction method for diabetic retinopathy - Google Patents

Abstract

The invention discloses an eye fundus image characteristics extraction method for diabetic retinopathy by utilizing an eye fundus image shot by a digital mydriatic-free eye fundus camera. The method comprises the following steps of (1) selecting an RGB channel of the eye fundus image; (2) positioning the optic disk of the eye fundus image; (3) extracting the hard exudate characteristics and the cotton wool spot characteristics of the eye fundus image, generating an extracted eye fundus image if at least one characteristic is discovered, if no characteristic is discovered, extracting microaneurysm characteristics and retinal hemorrhage characteristics, and generating an extracted eye fundus image. The method belongs to a non-intrusive technology, and is simple, rapid, convenient and effective, and the processed image is clear and obvious in symptom, so the diagnosis of a doctor is facilitated.

Description

The eye fundus image feature extracting method of DRP

Technical field

The present invention relates to digital image processing field, particularly a kind of eye fundus image feature extracting method of DRP.

Background technology

DRP (DiabeticRetinopathy, hereinafter to be referred as DR) is the highest complication of diabetic's incidence of disease, also maximum to eyesight influence, is that novel blind main cause appears in current 20-65 year adult.Clinically whether to occur that retinal neovascularization is as boundary, DR is divided into non-proliferative DRP (nonproliferativediabetic retinopathy, NPDR) (or claiming simple form or background type) and proliferative diabetic retinopathy (proliferativediabetic retinopathy, PDR), wherein at non-proliferation period, patient's retina there will be aneurysms (Microaneurysms, hereinafter to be referred as MAs), inter-retinal hemorrhage (Haemorrhages, hereinafter to be referred as Hs) and capillary seepage cause hard exudate (Hard exudates, hereinafter to be referred as EXs), propagation there will be cotton-wool patches (CottonWoolSpots in earlier stage, hereinafter to be referred as CWs).

Current diagnosis DR relies on manually to consult fundus photograph, the eye fundus image obtaining by fundus camera, this method is simple, image is easy to get, intuitively, be easy to preserve and record, and there is remarkable consistance and higher sensitivity with fundus fluorescein angiography diagnostic result, the advantages such as specificity, but the method relies on the visual inspection of oculist to eye fundus image substantially, this limitation of manually readding sheet method Existence dependency individual experience, and it is even that eye fundus image often has uneven illumination, the shortcomings such as blood vessel contrast is low, affect doctor's accurate judgement, therefore, how targetedly processing diabetic's eye fundus image, make Mas, Hs, EXs, the feature of CWs is clearer, it is this area problem demanding prompt solution always.

Summary of the invention

For present stage eye fundus image there is the shortcoming that uneven illumination is even, blood vessel contrast is low, utilizing numeral to exempt from mydriasis eye-ground photography technology combines with image processing and mode identification technology, a kind of DRP method for processing fundus images is provided, make eye fundus image feature more obvious, facilitate doctor to judge more accurately the feature of eyeground pathological changes.The present invention is achieved in that

An eye fundus image feature extracting method for DRP, comprises and utilizes the digital eye fundus image that mydriasis fundus camera is taken of exempting from, and it is characterized in that, takes following steps:

(1) eye fundus image RGB channel selecting;

(2) eye fundus image optic disk location;

(3) eye fundus image is carried out to hard exudate feature and cotton-wool patches feature extraction, if find at least one feature, generate the eye fundus image after extracting, if discovery feature not, carry out aneurysms feature and inter-retinal hemorrhage feature extraction, the eye fundus image after regeneration extracts.

Preferably, in the present invention, described step 1 eye fundus image RGB channel selecting refers to, R passage is selected in optic disk location, and G passage is selected in blood vessel segmentation, hard exudate feature extraction, cotton-wool patches feature extraction, aneurysms feature extraction and inter-retinal hemorrhage feature extraction.

Preferably, in the present invention, described step 2 eye fundus image optic disk location comprises:

(a) obtain optic disk candidate region based on Otsu Threshold segmentation;

(b) utilize the H passage in the HSV space of eye fundus image to extract the main blood vessel of retina and determine main vessel directions, obtain directional diagram; After in directional diagram, find out to the highest point of weighted registration filter response value, using this position as optic disk center; Finally, in the optic disk candidate region that utilizes described center to obtain, determine optic disk from step a.

Preferably, in the present invention, in described step 3, hard exudate feature and cotton-wool patches feature extracting method are, first obtain white image feature candidate region by eye fundus image, realize hard exudate feature and cotton-wool patches character separation by the svm classifier structure of two-layer cascade again, wherein one-level svm classifier architectural feature is the inside and outside color distortion in edges of regions intensity and region:

(a) edges of regions intensity:

$\mathrm{ES}=\sqrt{{\left(\frac{\∂f}{\∂x}\right)}^2+{\left(\frac{\∂f}{\∂y}\right)}^2}---\left(1\right)$

(b) the inside and outside color distortion in region:

$\mathrm{CD}=\frac{u_{\mathrm{inside}}}{u_{\mathrm{surrounding}}}---\left(2\right)$

Wherein, u represents the u passage of Luv color space.

Secondary svm classifier architectural feature is: (a) region area A; (b) u passage average μ in region _u; (c) v passage average μ in region _v.

Preferably, in the present invention, obtain white image feature candidate region and obtain by IFFCM method.

Preferably, in the present invention, in described step 3, aneurysms feature extraction comprises:

(a) blood vessel feature is removed, to eye fundus image G passage f _gcarry out gray scale morphology closed operation:

f _g1＝φ ^(sB)(f _g) (3)

Wherein φ represents gray scale morphology closed operation, and sB represents the morphosis element of size for s;

Then, utilize gray scale morphology corrosion to rebuild and fill f _gthe hole of middle existence:

$f_{g2}=R_{f_g}^{*}\left(f_m\right)---\left(4\right)$

Wherein,

r ^*represent gray scale morphology corrosion reconstruction, f _mfor rebuilding the marker in computing, f _gfor mask;

Finally, to f _g1with f _g2difference carry out Otsu Threshold segmentation and obtain blood vessel feature, that is:

f _g3＝T _Otsu(f _g1-f _g2) (5)

And utilize two-value morphology opening operation by f _g3in be less than or equal to aneurysms size region remove;

(b) aneurysms feature extraction

First, to f _gcarry out EMIN conversion, that is:

f _g4＝EMIN(f _g,t) (6)

F _g4for bianry image, t is the threshold value 0.05 of setting;

Then, from f _g4middle removal hard exudate feature and blood vessel feature, that is:

Wherein, ∧ represents that pointwise asks for minimum value, represents negate, f _eXsfor hard exudate feature, f _vesselfor blood vessel feature;

Finally, utilize the dimension information of MAs to obtain f in conjunction with two-value morphology opening operation _g5in aneurysms feature.

Preferably, in the present invention, in described step 3, the feature extracting method of inter-retinal hemorrhage is the local auto-adaptive region-growing method adopting based on multi-template matching, is specially:

(a) image is carried out to gamma correction inhomogeneous with removal of images gray scale, utilize the V passage V (i, j) in HSV space to calculate intensity correction values, this value is defined as:

$B_c(i,j)=\sqrt{1-{(V(i,j)-1)}^2}---\left(8\right)$

Then utilize contrast limited adaptive histogram equalization method to strengthen eye fundus image contrast;

(b) utilize template, as shown in Figure 1, identify unknown target image, obtain candidate region; Adopt Normalized Cross Correlation Function as similarity measure, be defined as:

$\mathrm{NCC}(i,j)=\frac{\underset{x=1}{\overset{X}{\mathrm{\Σ}}}\underset{y=1}{\overset{Y}{\mathrm{\Σ}}}{S}^{i,j}(x,y)T(x,y)}{\sqrt{\underset{x=1}{\overset{X}{\mathrm{\Σ}}}\underset{y=1}{\overset{Y}{\mathrm{\Σ}}}{\left[S^{i,j}(x,y)\right]}^2}\sqrt{\underset{x=1}{\overset{X}{\mathrm{\Σ}}}\underset{y=1}{\overset{Y}{\mathrm{\Σ}}}{\left[T(x,y)\right]}^2}}---\left(9\right)$

Wherein, r represents radius of a circle, and a represents the distance of circle to peripheral rectangle; T represents template, S ^i,jrepresent the search subgraph under template covering;

Average gray is introduced to above formula:

$\mathrm{NCC}(i,j)=\frac{\underset{x=1}{\overset{X}{\mathrm{\Σ}}}\underset{y=1}{\overset{Y}{\mathrm{\Σ}}}(S^{i,j}(x,y)-\stackrel{\‾}{S^{i,j}})(T(x,y)-\stackrel{\‾}{T})}{\sqrt{\underset{x=1}{\overset{X}{\mathrm{\Σ}}}\underset{y=1}{\overset{Y}{\mathrm{\Σ}}}{(S^{i,j}(x,y)-\stackrel{\‾}{S^{i,j}})}^2}\sqrt{\underset{x=1}{\overset{X}{\mathrm{\Σ}}}\underset{y=1}{\overset{Y}{\mathrm{\Σ}}}{(T(x,y)-\stackrel{\‾}{T})}^2}}---\left(10\right)$

represent respectively the average gray of search subgraph and template;

(c) centre of form of candidate region is defined as to Seed Points p; Next, calculate Seed Points and field pixel p around it _ibetween distance d, gray scale difference t; Then, according to select corresponding judgment threshold TR apart from d:

$\mathrm{TR}=\left\{\begin{array}{cc}\frac{|{\mathrm{KM}}_{M\&times;M}-{\mathrm{KM}}_{N\&times;N}|}{2},& 0<d\&le;D\\ \frac{|{\mathrm{KM}}_{M\&times;M}-{\mathrm{KM}}_{N\&times;N}|}{2}-d/t,& d>D\end{array}\right.---\left(11\right)$

If being less than or equal to corresponding threshold value, t this pixel is integrated into Seed Points region; Wherein KM _{m × M}, KM _{n × N}represent respectively pixel p _iand the gray average of Seed Points p region, D represents, apart from critical value, to work as p _iand when the distance between p is greater than D, reduce by 1 every t pixel judgment threshold TR, realize the extraction of the feature of inter-retinal hemorrhage.

The invention has the beneficial effects as follows, only need a number of units word to exempt from mydriasis fundus camera and just can exempt from mydriasis eye fundus image by what gather, can extract feature to eye fundus image according to hard exudate feature or cotton-wool patches feature or aneurysms feature or inter-retinal hemorrhage feature, carry out picture processing, the method belongs to a kind of non-intrusion type technology, have the advantages such as simple and easy, quick, effective, image symptom is after treatment clear, obvious, facilitates diagnosis.

Brief description of the drawings

Fig. 1 is related template and template matches process of inter-retinal hemorrhage (Hs) feature extraction;

Fig. 2 is feature extraction process flow diagram of the present invention;

Fig. 3 a is eye fundus image;

Fig. 3 b is optic disk positioning result;

Fig. 3 c is IFFCM cluster result;

Fig. 3 d is that MAs extracts result;

Fig. 3 e is NCC template matching results;

Fig. 3 f is eye fundus image after extracting.

Embodiment

Below in conjunction with specific embodiment, the present invention is illustrated, it should be pointed out that embodiment is interpreted as, for explaining content of the present invention but not limit the scope of the invention.

Embodiment 1

The eye fundus image that the present embodiment uses is to take by exempting from mydriasis fundus camera TopconNW100, as shown in Figure 3 a.(1) RGB channel selecting

The eyeground contained pigment of each physiological structure has different absorption characteristics, and different wave length monochromatic light is also different at the penetration performance on eyeground, according to the feature of the different focuses of DR, be divided into red image feature (MAs, Hs) and white image feature (EXs, CWs), for different detection targets, choose suitable color space representation form or subchannel according to its Spectral Characteristics Analysis result.

Optic disk is under 628nm ruddiness, and visibility is the highest.Optic disk edge clear under this wavelength light is shone, very poor from optic disk blood vessel visibility out, nerve fibre almost disappears, and optic disk is rendered as a reflection speck uniformly, therefore, for selecting R passage cutting apart of optic disk.And retina artery and vein trunk contrast is all higher in 478-589nm wavelength coverage, especially with under 570nm green glow, full visual field blood vessel visibility is best, vessel boundary is clear sharp keen, in the middle of the reflective dark background that is obviously positioned at blood post of axle, therefore, for selecting G passage cutting apart of blood vessel.For MAs and Hs, because itself and blood vessel have similar color, there is close spectral signature, therefore, its feature extraction also adopts G passage.EXs, CWs contrast in G passage is the highest, therefore also adopts G passage.

(2) eye fundus image optic disk location

Optic disk and EXs have stronger similarity aspect color and brightness.Given this, according to the direction of the main blood vessel of retina and converge on this characteristic of optic disk, the present invention utilizes the position of main blood vessel convergence point to judge the optic disk candidate region obtaining based on large Tianjin method (Otsu) Threshold segmentation, thereby realizes the method for optic disk Obtaining Accurate.

(a) the optic disk candidate region based on Otsu threshold method obtains

Obtain the region that shows as high brightness including optic disk, the i.e. candidate region of optic disk by the method;

(b) according to main vessel directions location optic disk

Be this characteristic of blood vessel convergence point according to optic disk, utilize the H passage in the HSV space of colored eye fundus image to extract the main blood vessel of retina and determine main vessel directions, obtain directional diagram; Then, find out the highest point of weighted registration filter response value in directional diagram, this position is as optic disk center; Finally, utilize " selecting " in the optic disk candidate region of this place-centric information from step a to go out real optic disk, as shown in Figure 3 b.

(3) hard exudate (EXs) that capillary seepage causes, the characteristic processing of cotton-wool patches (CWs)

FCM(FuzzyC-Means) algorithm is applicable to exist in image the occasion of uncertain and ambiguity, this feature detects particularly effective to CWs, but also there are some defects in this algorithm in actual applications, therefore select a kind of improved quick FCM(IFFCM) (see Shen Jianxin, Gao Weiwei. Chinese invention patent, " a kind of fuzzy clustering image partition method ". application number 201310072342.X, publication number 2013030700631260.) cut apart colored eye fundus image and obtain sugared net white image feature candidate region, as shown in Figure 3 c; In obtained sugared net white image feature candidate region, except EXs, CWs, also have some brilliant white background areas.For this three classification problem, adopt the SVM structure of two-layer cascade sort (all selecting radial basis kernel function), first utilize SVM by the sugared net white image character separation in candidate region out; SVM is by the EXs in white image feature for recycling, and CWs makes a distinction.For the differentiation of white image feature, the inside and outside color distortion of edges of regions intensity and region that is characterized as used.

(a) edges of regions intensity:

$\mathrm{ES}=\sqrt{{\left(\frac{\&PartialD;f}{\&PartialD;x}\right)}^2+{\left(\frac{\&PartialD;f}{\&PartialD;y}\right)}^2}---\left(1\right)$

(b) the inside and outside color distortion in region:

$\mathrm{CD}=\frac{u_{\mathrm{inside}}}{u_{\mathrm{surrounding}}}---\left(2\right)$

In formula, u represents the u passage of Luv color space.

For the classification of EXs and CWs, the feature of selecting is: (1) region area A; (2) u passage average μ in region _u; (3) v passage average μ in region _v.

(4) characteristic processing of aneurysms (MAs)

The isolated round dot of red or kermesinus that MAs is less than 125 μ m with diameter is present on retina, in eye fundus image G passage, show as that intra-zone has Continuous Gray Scale value but the strict higher isolated area of its external margin pixel value, therefore, can utilize EMIN conversion that its special this from eye fundus image G passage is extracted; But also exist and meet the region of this feature in EXs and blood vessel, therefore also need these two from the result EMIN conversion, to remove; Finally in the MAs candidate region obtaining, utilize the dimension information of MAs to obtain real MAs.Therefore, the feature extraction of MAs in eye fundus image mainly comprised to blood vessel feature extraction and realize the feature extraction of MAs according to dimension information.Specific algorithm is described below:

(a) blood vessel feature extraction.

Due to factors such as uneven illuminations, on the blood vessel in eye fundus image, there will be hole, blood vessel segment etc., this type of region also meets the characteristic that EMIN conversion institute can surveyed area, therefore, blood vessel be also must removal vacation sun.For the feature extraction of blood vessel, by the difference of two width images being carried out to the realization of Otsu Threshold segmentation, that is:

First, to eye fundus image G passage f _gcarry out gray scale morphology closed operation to eliminate blood vessel and correlative detail information, this morphological operation size of structure element used is greater than image medium vessels breadth extreme, that is:

f _g1＝φ ^(sB)(f _g) (3)

Wherein φ represents gray scale morphology closed operation, and sB represents the morphosis element of size for s.

Then, utilize gray scale morphology corrosion to rebuild and fill f _gthe hole of middle existence, for example red or kermesinus dot, and small holes on blood vessel, that is:

$f_{g2}=R_{f_g}^{*}\left(f_m\right)---\left(4\right)$

Wherein,

r ^*represent gray scale morphology corrosion reconstruction, f _mfor rebuilding the marker in computing, f _gfor mask.

Finally, to f _g1with f _g2difference carry out Otsu Threshold segmentation and obtain blood vessel candidate region, that is:

f _g3＝T _Otsu(f _g1-f _g2) (5)

And utilize two-value morphology opening operation by f _g3in be less than or equal to MAs size region remove.

(b) Mas feature extraction.Specifically describe as follows:

First, to f _gcarry out EMIN conversion, that is:

f _g4＝EMIN(f _g,t) (6)

F _g4for bianry image, t is setting threshold 0.05;

Then, from f _g4middle removal EXs feature and blood vessel feature, that is:

Wherein, ∧ represents that pointwise asks for minimum value, represents negate, f _eXsfor EXs testing result, f _vesselfor vessel segmentation.

Finally, utilize the dimension information of MAs to obtain f in conjunction with two-value morphology opening operation _g5in the characteristic information of Mas, thereby realize the feature extraction of Mas, as shown in Figure 3 d;

(5) feature extracting method of inter-retinal hemorrhage (Hs)

First, image is carried out to gamma correction inhomogeneous with removal of images gray scale.For this reason, utilize the V passage V (i, j) in HSV space to calculate intensity correction values, utilize this corrected value to carry out gamma correction to eye fundus image, this value is defined as:

$B_c(i,j)=\sqrt{1-{(V(i,j)-1)}^2}---\left(8\right)$

Then utilize contrast limited adaptive histogram equalization method (CLAHE) to strengthen eye fundus image contrast.Because Hs and blood vessel have similar spectral signature, therefore, some Hs that originally do not manifest in colored eye fundus image can display in pretreated G passage.

Template matches is to remove to identify unknown target image by known template, finally realizes the understanding to unknown images.For the Hs in eye fundus image, designed template and template matches process are shown in Fig. 1.Template T (r, a) (r represents radius of a circle, and a represents the distance of circle to peripheral rectangle, and by changing r, the value of a can be obtained different templates).Select normalized crosscorrelation (NCC) function as similarity measure, to be defined as herein:

$\mathrm{NCC}(i,j)=\frac{\underset{x=1}{\overset{X}{\mathrm{\&Sigma;}}}\underset{y=1}{\overset{Y}{\mathrm{\&Sigma;}}}{S}^{i,j}(x,y)T(x,y)}{\sqrt{\underset{x=1}{\overset{X}{\mathrm{\&Sigma;}}}\underset{y=1}{\overset{Y}{\mathrm{\&Sigma;}}}{\left[S^{i,j}(x,y)\right]}^2}\sqrt{\underset{x=1}{\overset{X}{\mathrm{\&Sigma;}}}\underset{y=1}{\overset{Y}{\mathrm{\&Sigma;}}}{\left[T(x,y)\right]}^2}}---\left(9\right)$

Wherein, T represents template, S ^i,jrepresent the search subgraph under template covering.

For overcoming the impact of illumination, average gray is introduced to above formula:

$\mathrm{NCC}(i,j)=\frac{\underset{x=1}{\overset{X}{\mathrm{\&Sigma;}}}\underset{y=1}{\overset{Y}{\mathrm{\&Sigma;}}}(S^{i,j}(x,y)-\stackrel{\&OverBar;}{S^{i,j}})(T(x,y)-\stackrel{\&OverBar;}{T})}{\sqrt{\underset{x=1}{\overset{X}{\mathrm{\&Sigma;}}}\underset{y=1}{\overset{Y}{\mathrm{\&Sigma;}}}{(S^{i,j}(x,y)-\stackrel{\&OverBar;}{S^{i,j}})}^2}\sqrt{\underset{x=1}{\overset{X}{\mathrm{\&Sigma;}}}\underset{y=1}{\overset{Y}{\mathrm{\&Sigma;}}}{(T(x,y)-\stackrel{\&OverBar;}{T})}^2}}---\left(10\right)$

represent respectively the average gray of search subgraph and template.

NCC template matching results as shown in Figure 3 e, for the Hs in eye fundus image, because its form size is irregular, therefore multiple templates that utilization designs are obtained its candidate region, thereby is determined the required seed of region growing.Template matching results is carried out to Threshold segmentation, get final product to obtain the candidate region of Hs.But due to factors such as uneven illumination are even, in optic disk and blood vessel, also there will be and meet the region that template characterizes, for this type of region is removed from acquired candidate region.

First, the centre of form of candidate region is defined as to Seed Points p; Next, calculate Seed Points and field pixel p around it _ibetween distance d, gray scale difference t; Then, according to select corresponding judgment threshold TR apart from d:

$\mathrm{TR}=\left\{\begin{array}{cc}\frac{|{\mathrm{KM}}_{M\&times;M}-{\mathrm{KM}}_{N\&times;N}|}{2},& 0<d\&le;D\\ \frac{|{\mathrm{KM}}_{M\&times;M}-{\mathrm{KM}}_{N\&times;N}|}{2}-d/t,& d>D\end{array}\right.---\left(11\right)$

If being less than or equal to corresponding threshold value, t this pixel is integrated into Seed Points region.Wherein KM _{m × M}, KM _{n × N}represent respectively pixel p _iand the gray average of Seed Points p region, D represents, apart from critical value, to work as p _iand when the distance between p is greater than D, reduce by 1 to prevent region outgrowth every t pixel judgment threshold TR.

To each candidate region, in predefined rectangular area, carry out above-mentioned local auto-adaptive region growing, thereby obtain its peripheral profile, finally realize the feature extraction to Hs.

After the extraction that the present embodiment obtains, eye fundus image is as shown in Fig. 3 f, and the image after this method is extracted, has got rid of the interfere informations such as optic disk blood vessel as seen, facilitates artificial observation.Fig. 2 is that the present embodiment extracts schematic flow sheet.

Claims (7)

1. an eye fundus image feature extracting method for DRP, comprises and utilizes the digital eye fundus image that mydriasis fundus camera is taken of exempting from, and it is characterized in that, takes following steps:

(1) eye fundus image RGB channel selecting;

(2) eye fundus image optic disk location;

(3) eye fundus image is carried out to hard exudate feature and cotton-wool patches feature extraction, if find at least one feature, generate the eye fundus image after extracting, if discovery feature not, carry out aneurysms feature and inter-retinal hemorrhage feature extraction, the eye fundus image after regeneration extracts.

2. the eye fundus image feature extracting method of DRP according to claim 1, it is characterized in that, described step 1 eye fundus image RGB channel selecting refers to, R passage is selected in optic disk location, and G passage is selected in blood vessel segmentation, hard exudate feature extraction, cotton-wool patches feature extraction, aneurysms feature extraction and inter-retinal hemorrhage feature extraction.

3. the eye fundus image feature extracting method of DRP according to claim 2, is characterized in that, described step 2 eye fundus image optic disk location comprises:

(a) obtain optic disk candidate region based on Otsu Threshold segmentation;

(b) utilize the H passage in the HSV space of eye fundus image to extract the main blood vessel of retina and determine main vessel directions, obtain directional diagram; After in directional diagram, find out to the highest point of weighted registration filter response value, using this position as optic disk center; Finally, in the optic disk candidate region that utilizes described center to obtain, determine optic disk from step a.

4. the eye fundus image feature extracting method of DRP according to claim 3, it is characterized in that, in described step 3, hard exudate feature and cotton-wool patches feature extracting method are, first obtain white image feature candidate region by eye fundus image, realize hard exudate feature and cotton-wool patches character separation by the svm classifier structure of two-layer cascade again, wherein one-level svm classifier architectural feature is the inside and outside color distortion in edges of regions intensity and region:

(a) edges of regions intensity:

$\mathrm{ES}=\sqrt{{\left(\frac{\&PartialD;f}{\&PartialD;x}\right)}^2+{\left(\frac{\&PartialD;f}{\&PartialD;y}\right)}^2}---\left(1\right)$

(b) the inside and outside color distortion in region:

$\mathrm{CD}=\frac{u_{\mathrm{inside}}}{u_{\mathrm{surrounding}}}---\left(2\right)$

Wherein, u represents the u passage of Luv color space.

Secondary svm classifier architectural feature is: (a) region area A; (b) u passage average μ in region _u; (c) v passage average μ in region _v.

5. the eye fundus image feature extracting method of DRP according to claim 4, is characterized in that, described white image feature candidate region obtains by IFFCM method.

6. the eye fundus image feature extracting method of DRP according to claim 5, is characterized in that, in described step 3, aneurysms feature extraction comprises:

(a) blood vessel feature is removed, to eye fundus image G passage f _gcarry out gray scale morphology closed operation:

f _g1＝φ ^(sB)(f _g) (3)

Wherein φ represents gray scale morphology closed operation, and sB represents the morphosis element of size for s;

Then, utilize gray scale morphology corrosion to rebuild and fill f _gthe hole of middle existence:

$f_{g2}=R_{f_g}^{*}\left(f_m\right)---\left(4\right)$

Wherein,

r ^*represent gray scale morphology corrosion reconstruction, f _mfor rebuilding the marker in computing, f _gfor mask;

Finally, to f _g1with f _g2difference carry out Otsu Threshold segmentation and obtain blood vessel feature, that is:

f _g3＝T _Otsu(f _g1-f _g2) (5)

And utilize two-value morphology opening operation by f _g3in be less than or equal to aneurysms size region remove;

(b) aneurysms feature extraction

First, to f _gcarry out EMIN conversion, that is:

f _g4＝EMIN(f _g,t) (6)

F _g4for bianry image, t is the threshold value 0.05 of setting;

Then, from f _g4middle removal hard exudate feature and blood vessel feature, that is:

Wherein, ∧ represents that pointwise asks for minimum value, represents negate, f _eXsfor hard exudate feature, f _vesselfor blood vessel feature;

Finally, utilize the dimension information of MAs to obtain f in conjunction with two-value morphology opening operation _g5in aneurysms feature.

7. the eye fundus image feature extracting method of DRP according to claim 5, is characterized in that, in described step 3, the feature extracting method of inter-retinal hemorrhage is the local auto-adaptive region-growing method adopting based on multi-template matching, is specially:

(a) image is carried out to gamma correction inhomogeneous with removal of images gray scale, utilize the V passage V (i, j) in HSV space to calculate intensity correction values, this value is defined as:

$B_c(i,j)=\sqrt{1-{(V(i,j)-1)}^2}---\left(8\right)$

Then utilize contrast limited adaptive histogram equalization method to strengthen eye fundus image contrast;

(b) utilize template (Fig. 1) to identify unknown target image, obtain candidate region; Adopt Normalized Cross Correlation Function as similarity measure, be defined as:

$\mathrm{NCC}(i,j)=\frac{\underset{x=1}{\overset{X}{\mathrm{\&Sigma;}}}\underset{y=1}{\overset{Y}{\mathrm{\&Sigma;}}}{S}^{i,j}(x,y)T(x,y)}{\sqrt{\underset{x=1}{\overset{X}{\mathrm{\&Sigma;}}}\underset{y=1}{\overset{Y}{\mathrm{\&Sigma;}}}{\left[S^{i,j}(x,y)\right]}^2}\sqrt{\underset{x=1}{\overset{X}{\mathrm{\&Sigma;}}}\underset{y=1}{\overset{Y}{\mathrm{\&Sigma;}}}{\left[T(x,y)\right]}^2}}---\left(9\right)$

Wherein, r represents radius of a circle, and a represents the distance of circle to peripheral rectangle; T represents template, S ^i,jrepresent the search subgraph under template covering;

Average gray is introduced to above formula:

$\mathrm{NCC}(i,j)=\frac{\underset{x=1}{\overset{X}{\mathrm{\&Sigma;}}}\underset{y=1}{\overset{Y}{\mathrm{\&Sigma;}}}(S^{i,j}(x,y)-\stackrel{\&OverBar;}{S^{i,j}})(T(x,y)-\stackrel{\&OverBar;}{T})}{\sqrt{\underset{x=1}{\overset{X}{\mathrm{\&Sigma;}}}\underset{y=1}{\overset{Y}{\mathrm{\&Sigma;}}}{(S^{i,j}(x,y)-\stackrel{\&OverBar;}{S^{i,j}})}^2}\sqrt{\underset{x=1}{\overset{X}{\mathrm{\&Sigma;}}}\underset{y=1}{\overset{Y}{\mathrm{\&Sigma;}}}{(T(x,y)-\stackrel{\&OverBar;}{T})}^2}}---\left(10\right)$

represent respectively the average gray of search subgraph and template;

(c) centre of form of candidate region is defined as to Seed Points p; Next, calculate Seed Points and field pixel p around it _ibetween distance d, gray scale difference t; Then, according to select corresponding judgment threshold TR apart from d:

$\mathrm{TR}=\left\{\begin{array}{cc}\frac{|{\mathrm{KM}}_{M\&times;M}-{\mathrm{KM}}_{N\&times;N}|}{2},& 0<d\&le;D\\ \frac{|{\mathrm{KM}}_{M\&times;M}-{\mathrm{KM}}_{N\&times;N}|}{2}-d/t,& d>D\end{array}\right.---\left(11\right)$

If being less than or equal to corresponding threshold value, t this pixel is integrated into Seed Points region; Wherein KM _{m × M}, KM _{n × N}represent respectively pixel p _iand the gray average of Seed Points p region, D represents, apart from critical value, to work as p _iand when the distance between p is greater than D, reduce by 1 every t pixel judgment threshold TR, realize the extraction of the feature of inter-retinal hemorrhage.

Images