CN105976385A - Image segmentation method based on image data field - Google Patents

Abstract

The invention relates to an image segmentation method based on an image data field. Conventional monochromatic image segmentation methods are generally based on the discontinuity and similarity of a pixel gray value; and a conventional segmentation algorithm is difficult to get the ideal segmentation effect under the condition of non-uniform illumination. Firstly, the input image is subjected to pre-processing filtering and then segmented by a watershed algorithm based on the eight-direction gradient, and then the average gray value of the edge point is obtained by the edge detection of the segmented image and is set as the seed value, and meanwhile, an image data filed is constructed, and the potential value function of the pixel point is calculated, and at the end, based on the seed value and the image data field, the feed filed automatic growth is carried out. The segmentation effect and the segmentation time of the image being nonuniform in illumination are greatly improved, and the demand of image identification can be met.

Description

A kind of image partition method based on view data field

Technical field

The invention belongs to image segmentation field, relate to the quick Accurate Segmentation method of the even image of a kind of uneven illumination.

Background technology

Image segmentation refers to that the similarity by characteristics of image or diversity segment the image into many sub-districts of continuous print Territory, and from subregion, extract part interested.In actual applications, these available features mainly include gray scale, Texture, edge, color etc., for different image processing requirements, have been proposed for thousands of kinds of processing methods at present, but do not have A kind of the most accurate general dividing method.

In engineer applied, image segmentation is the committed step realizing graphical analysis, is a weight of image processing system Wanting ingredient, have the highest practical value, be the follow-up basis carrying out graphical analysis, segmentation quality will directly affect last point The result of analysis, so the research carrying out improving the accuracy of image segmentation is necessary.

Summary of the invention

The present invention is directed to the deficiencies in the prior art, it is provided that a kind of quick Accurate Segmentation side for the even image of uneven illumination Method.

The present invention solves the technical scheme that technical problem taked:

Step (1). image input and Image semantic classification.

Step (2). use watershed algorithms based on eight direction gradients to carry out image segmentation.

Step (3). build view data field, calculate the gesture value function of each pixel.

Step (4). rim detection, calculate the average gray value of marginal point and be set to seed point value.

Step (5). integrating step (3), the data fields of (4) and seed points carry out automated seed region growing.

Step (6). homogeneous region merges output segmentation effect figure.

Beneficial effects of the present invention:

(1) have employed the watershed algorithm of eight direction gradients, take full advantage of the marginal information of image, decrease over-segmentation Degree.

(2) combine view data field and carry out region growing, it is contemplated that the basic structure of image itself, make segmentation effect more Accurately.

Accompanying drawing explanation

Fig. 1 is a kind of image partition method flow chart based on view data field.

Specific embodiment

Below in conjunction with Figure of description, the present invention is described further.

According to Figure of description (1), it is described in detail implementing step:

Step (1). image input and Image semantic classification.

In order to reach the effect of contrast, gather with piece image design sketch under different illumination conditions, enter the most respectively Row processes.Pretreatment to image, mainly includes image is carried out gray processing, filtering and noise reduction, histogram equalization, binaryzation etc. Several processing procedures.

Step (2). use watershed algorithms based on eight direction gradients to carry out image segmentation.

Watershed algorithm is a kind of image partition method based on gradient, the edge in image can be converted into " mountain Arteries and veins ", homogeneous area is converted in " mountain valley ", is broadly divided into sequence, floods two processes.

The first step: ask for image from all directions to gradient amplitude take maximum, gradient amplitude image is in the edge along object Having higher pixel value, and have relatively low pixel value elsewhere, watershed algorithm can produce watershed ridge along edge Line；Gradient calculation and angle differentiate, set point f (x, (use y) to represent pixel on gray level image by x, y) gray value at place Robinson gradient template and original image carry out convolution:

(x, y) at rate of change (namely f (x+Dx, the y+dy)-f of the direction for the directional derivative of a direction i.e. f (x, value y)), gradient direction i.e. the direction of directional derivative maximum when the direction is consistent with the direction of gradient, direction is led The value of number is equal to the mould of gradient.

Second step: sort, flood process.Assuming that definition territory F_fOn function f be pending image, its maximin It is respectively f_max=max (f), f_min=min (f).Set [F_f]^hThe upper limit threshold of representative function f, wherein [F_f]^h={ x ∈ F_f|f (x)≤f}。

Reg_Min_k(f)={ x ∈ F_f| x=Local minimum and f (x)=f}

The watershed region of F is setWithObtained by interative computation:

$X_{f_{\min }}={\[F\]}^{f_{\min }}$

$X_{f+1}=\mathrm{Re}g\_{\mathrm{Min}}_{f+1}\left(f\right)\∪{\mathrm{IZ}}_{F_f}\left(x_f\right)$

$\∀f\∈\[f_{\min },f_{\max }-1\]$

The watershed of function f isSupplementary set,Corresponding rise process, all of dam is corresponding function f's Watershed, produces many little receiving basins, causes image over-segmentation owing to the image of input exists too much very small region Seriously, so needing segmentation result is carried out similar area merging.

Step (3). build view data field, calculate the gesture value function of each pixel.

Using the gray scale difference value of image pixel and neighborhood territory pixel as the quality of data object, generate corresponding view data , image grain P (p, ε)={ q | q ∈ I ∧ | | p-q | |≤ε } is designated as the neighborhood territory pixel set of pixel centered by p.Common Optional potential function includes pseudo gravity, intends electrostatic field etc., and its Mathematical Morphology is respectively as follows:

ψ_x(y)=G*m_X/(1+(||x-y||/σ)^k)；

ψ x (y)=m_X/(4πε₀*(1+(||x-y||/σ)^k))；

Wherein m_X>=0 is data object quality, represents field source intensity；, G is that constant acceleration of gravity | | x-y | | represents two The spacing of person, natural number k is range index, and σ ∈ (0 ,+∞) is referred to as factor of influence.

Gradient according to potential function is the field intensity function in the corresponding field of force, and the field intensity vector of data particle x is:

$F_x\left(t\right)=\▿{\mathrm{\ψ}}_x\left(t\right)=\frac{2}{{\mathrm{\σ}}^2}{\mathrm{\Σ}}_{y\∈\mathrm{\Ω}}\left(p_y\right(t)-p_x(t\left)\right)*m_y\left(t\right)*\exp \left(\frac{-\left|\right|p_y\left(t\right)-p_x\left(t\right)|{|}^2}{{\mathrm{\σ}}^2}\right)$

Wherein | | p_y(t)-p_x(t) | | represent the distance between x and y, quality m_yT () regards the action intensity of field source, σ as > 0 is factor of influence, determines the coverage of particle interaction force sometime.

Step (4). rim detection, calculate the average gray value of marginal point and be set to seed point value.

The method that rim detection is the most general is the discontinuity of detection brightness value, and this discontinuity can be by the one of image Rank and second dervative detection；Here use Canny operator to perform rim detection, first carry out gaussian filtering and eliminate noise, then Calculate the gradient at pixelDirection α (x, y)=arctan (G with edge_y/G_x)；Then paint The grey level histogram of marginal point pixel processed, takes its meansigma methods and is labeled as seed point value.

Step (5). data fields and seed points that integrating step (3), (4) obtain carry out automated seed region growing.

Region growing is the image partition method of a kind of serial region segmentation.It is from certain pixel, according to certain Criterion, is gradually added neighborhood pixels, and when certain conditions are met, region growing terminates.The quality of region growing is decided by 1. Choosing of initial point (seed points)；2. growth criterion；3. end condition.Seed points pixel value step (4) obtained is as life Long initial point, using the gesture value function of step (3) as growth criterion, and chooses suitable growing threshold and carries out region growing.Tool Body step is as follows:

One, the average gray value that step (4) is obtained as growth initial value, find the 1st the most do not belong to initial Value point, if this pixel is (x0, y0).

Two, centered by (x0, y0), it is contemplated that (x, if y) (x0, y0) meets growth criterion to 4 neighborhood territory pixels of (x0, y0)

X (x-1, y)+X (x+1, y)+X (x, y-1)+X (x, y+1) > V, V=255,

(x, y) represents the pixel value during labelling without Seed Sequences of point (x0, y0) to X, if having, is 0.Then will (x, y) with (x0, y0) merges (in the same area), simultaneously will (x, y) press-in storehouse.

Three, from storehouse, take out a pixel, it is returned to step 2 as (x0, y0).

Four, when storehouse is for returning to step one time empty.

Five, when repetition step one to four has ownership until each point in image, growth terminates.

Step (6). homogeneous region merges output segmentation effect figure.

Region merging technique, obtains the zonule P of many, it is assumed that its barycenter is (x after step (5) region growing₀,y₀):

$x_0=\frac{\underset{(x,y)\∈P}{\mathrm{\Σ}}xX(x,y)}{\underset{(x,y)\∈P}{\mathrm{\Σ}}X(x,y)},y_0=\frac{\underset{(x,y)\∈P}{\mathrm{\Σ}}yX(x,y)}{\underset{(x,y)\∈P}{\mathrm{\Σ}}X(x,y)}.$

X in formula (x, y) is the gray value of this point, if p^*∈ P is the center pixel of any one subregion after growth, ψ_ε(p^*)≤α represents its gesture value, and α is growing threshold, forIf meeting condition (a) q ∈ ξ (p1^*, ε) and ∧ ψ_q(p1^*)≤β；B () works as k > 1, Then q and p^*Constituting homogeneity to attract, wherein ε is the factor of influence of gesture value function, and β is predefined gesture value threshold value.

Claims (3)

1. an image partition method based on view data field, it is characterised in that comprising the concrete steps that of the method:

Step (1). image input and Image semantic classification；

Step (2). use watershed algorithms based on eight direction gradients to carry out image segmentation；

Step (3). build view data field, calculate the gesture value function of each pixel；

Step (4). rim detection, calculate the average gray value of marginal point and be set to seed point value；

Step (5). integrating step (3), the data fields of (4) and seed points carry out automated seed region growing；

Step (6). homogeneous region merges output segmentation effect figure.

A kind of image partition method based on view data field the most according to claim 1, it is characterised in that: step (2) Specifically:

The first step: ask for image from all directions to gradient amplitude take maximum；Gradient calculation and angle differentiate, set point f (x, y) generation Pixel on table gray level image (x, y) gray value at place, carry out convolution with Robinson gradient template and original image:

(x, y) at the rate of change of the direction, when the direction is consistent with the direction of gradient for the directional derivative of a direction i.e. f Time, the value of directional derivative is equal to the mould of gradient；

Second step: sort, flood process；Assuming that definition territory F_fOn function f be pending image, its maximin is respectively For f_max=max (f), f_min=min (f)；Set [F_f]^hThe upper limit threshold of representative function f, wherein [F_f]^h={ x ∈ F_f|f(x)≤ f}；

Reg_Min_k(f)={ x ∈ F_f| x=Local minimum and f (x)=f}

The watershed region of F is setWithObtained by interative computation:

$X_{f_{\min }}={\left[F\right]}^{f_{\min }}$

$X_{f+1}=\mathrm{Re}g\_{\mathrm{Min}}_{f+1}\left(f\right)\∪{\mathrm{IZ}}_{F_f}\left(x_f\right)$

$\∀f\∈\[f_{min},f_{\max }-1\]$

The watershed of function f isSupplementary set,Corresponding rise process, point water of all of dam is corresponding function f Ridge.

A kind of image partition method based on view data field the most according to claim 1, it is characterised in that: step (5) Specifically:

The first step, the average gray value that step (4) is obtained as growth initial value, find the 1st the most do not belong to initial Value point, if this pixel is (x0, y0)；

Second step, centered by (x0, y0), it is contemplated that (x, if y) (x0, y0) meets growth criterion for 4 neighborhood territory pixels of (x0, y0)

X (x-1, y)+X (x+1, y)+X (x, y-1)+X (x, y+1) > V, V=255,

(x, y) represents the pixel value during labelling without Seed Sequences of point (x0, y0) to X, if having, is 0；Then will (x, y) with (x0, Y0) it is incorporated in the same area, simultaneously will (x, y) press-in storehouse；

3rd step, from storehouse take out a pixel, it is returned to second step as (x0, y0)；

4th step, when storehouse is for returning to the first step time empty；

When 5th step, first to fourth step that repeats have ownership until each point in image, growth terminates.