CN102567731A - Extraction method for region of interest - Google Patents

Abstract

The invention discloses an extraction method for a region of interest. The extraction method specifically comprises the following steps of: 1, preprocessing an original image; 2, calculating a salient map for the original image; 3, cutting the original image by using a watershed segmentation algorithm; 4, calculating interest of each region according to the salient map and a segmentation result; and 5, detecting the region of interest. According to the extraction method, the region of interest in the image can be automatically searched and found; an Itti model for simulating a human attention mechanism is adopted; and the found region of interest is basically matched with the subjective feeling of a person. Compared with manual marking of the region of interest, the extraction method is faster and more accurate.

Description

A kind of area-of-interest exacting method

Technical field

The invention belongs to the Computer Image Processing field, proposed a kind of area-of-interest exacting method.

Background technology

Digital picture and rapid development of Internet, the explosive growth that has brought the digital picture resource, because it is the expressiveness of image is stronger than literal, also more directly perceived.But with respect to the expression-form of literal, image generally all takies a large amount of storage spaces, and how expressing the content that image will express with storage space still less just becomes a vital problem.Consider that picture material is rich and varied; And user's concern often is the some key areas in the picture; The JPEG2000 standard proposes a kind of encoding region of interest mechanism, and promptly picture can be divided into a plurality of zones, adopts the different compression rate to encode to different zones.This mechanism can address the above problem to a certain extent, but also has a problem not solve, and promptly how to find area-of-interest.The JPEG2000 standard needs the artificial mark of user area-of-interest, and this makes the application of method that significant limitation arranged, and seldom has the user to use this mechanism in the standard, because operate cumbersome.If the imagination method can detect the area-of-interest in the image automatically, perhaps only need user interactions seldom can find the area-of-interest in the image, this mechanism will inevitably receive more favor.

Find area-of-interest, at first will each Region Segmentation in the picture be come out, this will use image segmentation algorithm.At present; Image segmentation method commonly used has similarity measure algorithm, interest tolerance detection technique and from shallow depth image, extracts the algorithm etc. of area-of-interest automatically; The similarity measure algorithm is unusual difficulty when being the new feature of reflection high-level semantic in extracting image; Key is to have huge wide gap between image low-level feature and the high-level semantic, and the area-of-interest that utilizes flex point to extract image has significant limitation.The shortcoming of flex point detecting device is because flex point concentrates on texture region, so the point-of-interest that extracts with the viewpoint detecting device is concentrated the zone that is distributed in more than the texture, and at the very coefficient of the few areal distribution of texture.The too uneven distribution of this texture is unfavorable for intactly describing the content of image each several part.

Summary of the invention

The present invention is in order to solve the limitation of JPEG2000 standard encoding region of interest Technology Need manual markings area-of-interest; A kind of method of extracting interesting image regions has been proposed; Adopt the mechanism of watershed segmentation algorithm combination Itti attention model algorithm, realize the automatic detection of interesting image regions.The watershed segmentation algorithm is a kind of based on morphologic partitioning algorithm, and each zone in the image that can efficiently and accurately is cut apart.And the Itti attention model has been simulated biological vision noticing mechanism characteristic; Select for use the not good enough sample mode of central peripheral to extract characteristics of image; The characteristics of image of different dimensions is fused to remarkable figure; The focus-of-attention that the competition of significant point process obtains merges remarkable figure and watershed segmentation zone and obtains area-of-interest as the seed points of watershed segmentation.

A kind of area-of-interest exacting method of the present invention specifically may further comprise the steps:

Step 1: original image is carried out pre-service.

Step 2: original image is calculated significantly figure.

Step 3: original image is cut apart with the watershed segmentation algorithm.

Step 4:, calculate the degree interested in each zone according to remarkable figure and segmentation result.

Step 5: carry out area-of-interest and detect.

Advantage of the present invention and good effect are:

(1) the present invention can seek the area-of-interest in the positioning image automatically, has adopted the Itti model of simulating human notice mechanism, and area-of-interest that finds and people's subjective sensation are identical basically.With respect to handmarking's area-of-interest, the present invention is also more accurate more fast.

(2) the inventive method is supported the area-of-interest of arbitrary shape.Jpeg2000 at first puts forward the notion of area-of-interest, but from the consideration of compressing, Jpeg2000 only supports rule zones such as circle, rectangle.By contrast, the inventive method is supported any regular or irregular area-of-interest according to segmentation result.For example the profile of people's face in the image if identify area-of-interest with regular circular or rectangle, has bigger error.

(3) the present invention can be used as the pre-treatment step of some image processing algorithms, in ROI Image Coding, owing to can find the area-of-interest in the image, therefore, can encode respectively to area-of-interest and background area very easily.In popular CBIR (Content-Based Image Retrieval) system, because area-of-interest more can be represented the content of entire image, therefore, can give bigger weights to the characteristic of area-of-interest, like this, help improving whole accuracy rate.

Description of drawings

Fig. 1 is the whole flow chart of steps of extraction area-of-interest method of the present invention;

Fig. 2 is a comparison diagram before and after sample picture one pre-service;

Fig. 3 is a comparison diagram before and after sample picture two pre-service;

Fig. 4 is a comparison diagram before and after sample picture three pre-service;

Fig. 5 is a comparison diagram before and after sample picture four pre-service;

Fig. 6 is the remarkable figure of sample picture one;

Fig. 7 is the remarkable figure of sample picture two;

Fig. 8 is the remarkable figure of sample picture three;

Fig. 9 is the remarkable figure of sample picture four;

The comparison diagram as a result that Figure 10 adopts watershed segmentation algorithm and other partitioning algorithms commonly used to cut apart for sample picture one;

The comparison diagram as a result that Figure 11 adopts watershed segmentation algorithm and other partitioning algorithms commonly used to cut apart for sample picture two;

The comparison diagram as a result that Figure 12 adopts watershed segmentation algorithm and other partitioning algorithms commonly used to cut apart for sample picture three;

The comparison diagram as a result that Figure 13 adopts watershed segmentation algorithm and other partitioning algorithms commonly used to cut apart for sample picture four;

Figure 14 is the area-of-interest mask that sample picture one utilization this method is extracted automatically;

Figure 15 is the area-of-interest mask that sample picture two utilization this method are extracted automatically;

Figure 16 is the area-of-interest mask that sample picture three utilization this method are extracted automatically;

Figure 17 is the area-of-interest mask that sample picture four utilization this method are extracted automatically.

Embodiment

To combine accompanying drawing and instance that the present invention is done further detailed description below.

The present invention has provided a kind of area-of-interest exacting method based on watershed segmentation (Segmentation Method Based On Region Of Interest; SM-Roi); Method is according to the vision noticing mechanism characteristic of biology; Select for use the not good enough sample mode of central peripheral to extract characteristics of image, the characteristics of image of different dimensions is fused to remarkable figure; The focus-of-attention that the competition of significant point process obtains merges remarkable figure and watershed segmentation zone and obtains area-of-interest as the seed points of watershed segmentation; Follow to return and suppress and contiguous preferential accurate selection and shifting attention focus; Thereby obtain the significance level or the degree interested in zone; Through experiment; The result shows that this method meets biological vision noticing mechanism, when detecting area-of-interest automatically, can effectively reduce too and cut apart, and also can handle big object preferably.

Method among the present invention is based on image segmentation, and strictly speaking, this method also is a kind of dividing method.With respect to additive method; This method has added the step of graphical analysis, through utilization Itti visual attention model each sub regions is analyzed, and method can navigate to the area-of-interest in the image preferably; As shown in Figure 1, the concrete performing step of the inventive method is following:

Step 1: original image is carried out pre-service.

The gauge point that employing is come by the significant point expansion is as the basis for estimation of area-of-interest, and main departure degree by pixel attributive analysis, regional mean value computation and bank section pixel thereof expands pixel 3 parts and forms.Because each pixel is different with the departure degree of image average in the image; Set out a parameter K through emulation experiment; As long as the pixel value of pixel is less than parameter; Just strengthen at this pixel automatically, Filtering Processing, color space conversion, threshold value, area-of-interest are set extract and image restoration, concrete pre-treatment step is following:

Step 1.1: calculate each pixel in the image (x, pixel value s y) _i, wherein, i is the pixel subscript.

Step 1.2: the number of supposing the pixel in the image is N, calculates the pixel average m of entire image:

$m=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{s}_i---\left(1\right)$

Wherein: total number of pixels of N presentation video, s _iRepresent i gray values of pixel points.

Step 1.3: according to average m and pixel (x, pixel value s y) _i, the departure degree of each pixel and its average during equilibrium is published picture and looked like:

w _i＝‖s _i-m‖ ²(2)

Wherein: m is the mean value that step 1.2 is obtained.

Step 1.4: setup parameter K is 1/10 of an image average, as the departure degree w of image pixel value and mean value _iDuring less than K, just on the pixel value of its neighborhood, add K; If departure degree w _iMore than or equal to K, the pixel value of its neighborhood point equals the pixel value of pixel;

Step 2: original image is calculated significantly figure.

At first pretreated image is carried out filtering.Any image all can have noise and distortion, and the present invention adopts medium filtering to avoid the fuzzy of image detail, protects the image border simultaneously, the filtering random impulsive noise, and filtered image outline is more clear.What the present invention adopted is median filtering method; It is roamed a template pixel that contains odd number earlier in the drawings; And its template center overlapped with certain location of pixels in the image; Read each the respective pixel gray-scale value in the template, then these gray-scale values are sorted from small to large, find out its intermediate value and give the pixel of the center of corresponding templates it.

Carry out color space conversion then.Because the HIS color space more meets people's normal vision observation property; It is to describe color characteristic with tone saturation degree and brightness thereof; Have the visuality of radical sign than RGB color space, thus the image after the present invention will handle by the RGB color space conversion to corresponding HIS color space.

$\left\{\begin{array}{c}H=1/360[90-\arctan (F/3)+\{0,G>B;180,G>B\left\}\right]\\ S=1-[\min (R,G,B)/I]\\ I=(R+G+B)/3\\ F=(2R-G-B)/(G-B)\end{array}---\left(3\right)\right.$

Wherein: G representes the value of Green in the rgb space (green) component; B representes the value of Blue in the rgb space (blueness) component, and R representes the value of Red in the rgb space (redness) component, and H representes the value of Hue in the HIS space (tone) component; I representes the value of Intensity in the HIS space (intensity) component; S representes the value of Saturation in the HIS space (saturation degree), and F is an intermediate value of asking for convenience of calculation.

Certain regional conspicuousness is not to embody through this regional pixel value in the image, but contrast through this zone and peripheral region embodies, and the zone that has high-contrast in the image is easy to obtain people's attention.Therefore, just can obtain each pixel conspicuousness through each the pixel global contrast in the computer picture.Provided in the prior art and calculated the method for contrast, but implemented all more complicated.Therefore the present invention adopts a kind of method of simple and effective calculating contrast, and it is following that the tone overall situation correlative value of calculating pixel x gets formula:

$S_h\left(x\right)=\underset{V_h=V_h\min }{\overset{V_h\max }{\mathrm{\Σ}}}\left(\right|H\left(x\right)-V_h|\·{\mathrm{hist}}_h\left(v_h\right))---\left(4\right)$

$S_s\left(x\right)=\underset{V_s=V_s\min }{\overset{V_s\max }{\mathrm{\Σ}}}\left(\right|\mathrm{Sue}\left(x\right)-V_s|\·{\mathrm{hist}}_s\left(v_s\right))---\left(5\right)$

$S_i\left(x\right)=\underset{V_i=V_i\min }{\overset{V_i\max }{\mathrm{\Σ}}}\left(\right|I\left(x\right)-V_i|\·{\mathrm{hist}}_i\left(v_i\right))---\left(6\right)$

$S_x=\sqrt{S_h{\left(x\right)}^2+S_s{\left(x\right)}^2+S_i{\left(x\right)}^2}---\left(7\right)$

Wherein: in the formula (4), S _h(x) significance of the h component of remarked pixel x, V _hThe minimum value of h component in the min presentation video, V _hThe maximal value of the h component of max remarked pixel x, hist _hThe statistics with histogram result of presentation video h component, in the formula (5), S _s(x) significance of the s component of remarked pixel x, V _sThe minimum value of s component in the min presentation video, V _sThe maximal value of the s component of max remarked pixel x, hist _sThe statistics with histogram result of presentation video s component, in the formula (6), S _i(x) significance of the h component of remarked pixel x, V _iThe minimum value of i component in the min presentation video, V _iThe maximal value of the i component of max remarked pixel x, hist _iThe statistics with histogram result of presentation video i component.In the formula (7),, utilize formula (7) just can estimate each pixel conspicuousness in the image with the synthesis result significance as a whole of above-mentioned three component significances.

Step 3: original image is cut apart with the watershed segmentation algorithm.

The notion of watershed divide is treated to the basis image is carried out three-dimensional visualization: wherein two is coordinate, and another is a gray level.For the explanation of a kind of like this " topography ", consider three types of points: the point that (a) belongs to the locality minimum value; (b) when a water was placed on certain point locational, water was bound to drop to a single minimum point; (c) when water was on the position of certain point, water can flow to more than such minimum point equiprobably.To a specific regional minimum value, the point of satisfy condition (b) set is called " catchment basin " or " watershed divide " of this minimum value.The point point set of (c) of satisfying condition is set up into the crest line of topographical surface jointly, and term is called " cut-off rule " or " watershed line ".

Fundamental purpose based on the partitioning algorithm of these notions is to find out watershed line.Basic thought is very simple: suppose on the position of each regional minimum value, to make a call to a hole and whom lets gush out from the hole with uniform climbing speed, flood whole landform from low to high.When the different water that converge in the basin will condense together, the dam of building was with organized polymerization originally.The top that water can only arrive dam is in the degree on the waterline.The border of these dams is corresponding to the cut-off rule of watershed divide.So they extract (continuous) separatrix by watershed algorithm.

The detailed process of original image being cut apart with the watershed segmentation algorithm is following:

Make M ₁, M ₂... M _R(this is the typical gradient image of a width of cloth for x, the coordinate set of local minimum point y) for presentation video g.Make C (M _i) be the set of the coordinate of a point, these points are positioned at and local minimum M _iIn the catchment basin that (recall in the catchment basin whichsoever place all form UNICOM's component) interrelates.Symbol min and max represent g (x, minimum value y) and maximal value.At last, make T [n] denotation coordination (s, set t), wherein g (s, t)＜n, that is:

On how much, T [n] be g (x, the coordinate set of the point in y), the point in the set all be positioned at plane g (x, y)=below of n.

Along with water level from n=min+1 to n=max+1, constantly to increase, the landform in the image can be covered by water.Cover each stage in the process of landform at water level, method need know that all the point that is under the water level counts out.From conceptive, suppose coordinate among the T [n] be in g (x, y)=the n plane under, and be marked as black, so other coordinate is marked as white.So, when when water level increases with any increment n, observe the xy plane from the top down, can see a width of cloth bianry image.In image black color dots corresponding to be lower than in the function plane g (x, y)=point of n.

Make C _n(M _i) coordinate set of expression catchment basin mid point.This basin is with relevant in covered minimum value of n stage.C _n(M _i) also can be counted as the bianry image that provides by following formula:

C _n(M _i)＝C(M _i)∩T[n] (9)

Wherein, C (M _i) and T [n] represent the set of two coordinate points respectively, definition is with reference to formula (8), C _n(M _i) represent this two intersection of sets collection.In other words, if (x, y) ∈ C (M _i) and ((x y) has C then in the position for x, y) ∈ T [n] _n(M _i)=1, otherwise, C _n(M _i)=0.Explanation on result's geometry is very simple hereto.Only needing to use " with (AND) " operator that the bianry image among the T [n] is separated in n the stage that water overflows gets final product.T [n] is and local minimum M _iThe set that interrelates.

Next, make C [n] be illustrated in the intersection of the part that n stage catchment basin do not had by water logging:

$C\left[n\right]=\underset{j=1}{\overset{R}{\∪}}{C}_n\left(M_j\right)---\left(10\right)$

Wherein, C _n(M _j) definition with reference to formula (9), R is total number of local minimum.Make C [max+1] be the intersection of all catchment basins then:

$C[\max +1]=\underset{j=1}{\overset{R}{\∪}}C\left(M_j\right)---\left(11\right)$

Wherein, max is g (x, maximal value y), the max pixel value that promptly occurs in the image.Can find out and be in C _n(M _i) and T [n] in element term of execution of algorithm, can not be replaced, and the element number in these two set and n keep simultaneous growth.Therefore, C [n-1] is the subclass of set C [n].According to formula (10) and (11), C [n] is the subclass of T [n], so C [n-1] is the subclass of T [n].From this conclusion, each the UNICOM's component that draws among important result: the C [n-1] all is UNICOM's component of T [n] just.

When beginning, the algorithm of looking for watershed line sets C [min+1]=T [min+1].Algorithm gets into recursive call then, supposes when n goes on foot, and has constructed C [n-1].The process of trying to achieve C [n] according to C [n-1] is following: make Q represent the set of UNICOM's component among the T [n].Then, for each component q ∈ Q of UNICOM [n], following three kinds of possibilities are arranged:

(a) q ∩ C [n-1] is empty.

(b) q ∩ C [n-1] comprises UNICOM's component among the C [n-1].

(c) q ∩ C [n-1] comprises unnecessary one the UNICOM's component of C [n-1].

Depend on this 3 conditions according to C [n-1] structure C [n].When running into a up-to-date minimum value eligible (a), then incorporate q into C [n-1] and constitute C [n].When q was arranged in the catchment basin of some local minimum formation, eligible (b) integrated with C [n-1] with q and constitute C [n] this moment.When running into the topographical crest of two or more catchment basins of all or part of separation, eligible (c).Further the water filling meeting causes the water in different basins to condense together, thereby water level is reached unanimity.Therefore, must in q, set up a dam (if relate to a plurality of basins will set up do dam) overflows to stop intrabasinal water more.

Through use with g (x, the corresponding n value of gray-scale value that exists in y) can be improved efficiency of algorithm; (x, histogram y) can be confirmed these values and minimum value and maximal value according to g.

Step 4:, calculate the degree interested in each zone according to remarkable figure and segmentation result.

The size that at first will significantly scheme is shrunk to the size the same with original image in proportion, and like this, significantly the value of each point among the figure can correspond to the degree interested of pixel in the original image, and the size of value is all between 0 and 1.Since not of uniform size the causing in each zone, if simply add up each regional value, will certainly be with the area maximum region as area-of-interest, and the maximum zone of area sometimes possibly be background area etc.Certainly, can get the mean value in zone, but may cause like this choosing some because the tiny area that over-segmentation causes does not square with the fact, therefore, the present invention has taken all factors into consideration two kinds of factors, and has provided the computing formula of a kind of zone degree interested:

$\mathrm{Interest}\left(R_i\right)={\mathrm{\γ}}_1\×\frac{\mathrm{Num}\left(R_i\right)}{\mathrm{Num}\left(R\right)}+{\mathrm{\γ}}_2\×\frac{\underset{j=1}{\overset{\mathrm{Num}\left(R_i\right)}{\mathrm{\Σ}}}{r}_j}{\mathrm{Num}\left(R_i\right)}---\left(12\right)$

Wherein, γ ₁+ γ ₂=1, represent the weights coefficient of region area and average degree interested, Num (R respectively _i) and Num (R) the pixel number of being with not regional i and overall region to comprise respectively, r _jCertain any degree interested in the expression zone.

Step 5: carry out area-of-interest and detect.

The area-of-interest detection technique of the inventive method comprises image segmentation and two parts of interest tolerance, and region detecting part is divided with gradient watershed transform split image, and (x y) has described the object grey scale change to gradient G, is expressed as

$G(x,y)=\sqrt{{(I(x,y)*G_x)}^2+{(I(x,y)*G_y)}^2}---\left(13\right)$

Wherein: (x y) is gray-scale map, G to I _x, G _yBe the sobel edge mask and

At first image segmentation is become several regions, then gradient image is carried out the watershed divide mark and produce mark matrix L rgb, select to note that with focus-of-attention the zone produces mask mask figure at last according to image gradient.

Interest tolerance part; Adopt vision noticing mechanism to build interest-degree, choose the watershed segmentation result, use the Itti attention model to calculate the interest-degree in focus-of-attention tolerance zone; The conspicuousness in zone depends on the conspicuousness of graphic feature; As: brightness, color and direction etc., characteristic remarkable property obtains through the difference of gaussian sampling of computed image regional center and periphery, and formula is:

$\mathrm{DOG}(x,y)=\frac{1}{2{\mathrm{\π\σ}}_c^2}\exp |-\frac{x^2+y^2}{{2\mathrm{\σ}}_c^2}|-\frac{1}{2{\mathrm{\π\σ}}_s^2}\exp |-\frac{x^2+y^2}{{2\mathrm{\σ}}_s^2}|---\left(14\right)$

Wherein, (formula (14) is the expression of a double gauss differential mode type on the whole, σ for x, y) the pixel coordinate of presentation video _cAnd σ _sBe respectively the parameter of model, need to confirm, go 0.5 and 1.5 in this method respectively through experiment.Respectively to brightness I, color characteristic figure C and direction character figure O ask the conspicuousness of local component with formula (14), and be last, and whole remarkable figure S is the combination of brightness figure I, color characteristic figure C and direction character figure O, and formula is:

S＝w _i×N(I)+w _c×N(C)+w _o×N(O) (15)

In the formula, N (.) is a normalized factor, is about to a component value and constrains in [0,1] interval w _i, w _cAnd w _oBe respectively the characteristic weights of each component, and w _i+ w _c+ w _o=1.

After significantly being schemed S, next step need find the focus-of-attention among the remarkable figure, and focus-of-attention is the strongest point of vision significance, and being needs the at first place of concern in the scene, and focus-of-attention has selectivity and metastatic characteristics.The selection of focus-of-attention is to obtain through the neural net method that the victor is a king with shifting, and formula is:

$V(t+\mathrm{\δt})=[1-\frac{\mathrm{\δt}}{\mathrm{CR}}]V\left(t\right)+\frac{\mathrm{\δt}}{C}I\left(t\right)---\left(16\right)$

In the formula, C representes electric capacity, and R is a resistance, and V is a mode voltage, and following formula is represented the output voltage V (t) in the known t moment and the mode voltage V (t+ δ t) that input current I (t) produces behind time δ t, through after a while, and the voltage integrating meter granting of generation.When choosing focus-of-attention to image, with the integration granting neuron array of significantly scheming to regard as a two dimension, the value of pixel among the corresponding significantly figure of neuronic input current; Each neuronic film in will significantly scheming is then led through electricity and is converted neuronic input current in the WTA network to.The WTA network also be also should two dimension integration provide neuron town Liu, it is littler than neuronic time constant among the remarkable figure to play neuronic time constant, current potential rises to faster than the neuron among the remarkable figure.Like this, the neuron in the WTA network always produces earlier than the neuron among the remarkable figure and provides, and the neuron that generation is at first provided is just corresponding to significantly significantly being worth maximum neuron, i.e. focus-of-attention among the figure.

Before focus-of-attention shifts, need choose next focus-of-attention, that is to say that focus-of-attention shifts a capital and produces a focus-of-attention, the conspicuousness of the marking area of having selected can be suppressed, and is called to return inhibition.For this reason, return inhibition with a width of cloth and scheme to suppress current region, make notice turn to next zone; When carrying out the k time focus-of-attention transfer, return inhibition figure IR at the k width of cloth ^kIn, will belong to interior ground of (k-1) individual marking area pixel value and all be changed to 0, the pixel value of remaining position returns inhibition figure IR with respect to (k-1) width of cloth ^K-1Constant, as shown in the formula:

IR ⁰(x，y)＝1

In the formula, R ^kBe k marking area: (x y) is coordinate points; IR ^k(x, when y) k focus-of-attention chosen in expression, the k width of cloth returned coordinate points among the inhibition figure (x, value y).The area-of-interest of the public definite image of degree regional interested that calculates according to the focus-of-attention that finds and step 4 at last.

Embodiment:

The present invention realizes the extraction of area-of-interest automatically, and handling for follow-up low layer encoding operation or high-level semantic provides good precondition.Extremely shown in Figure 5 like Fig. 2; Be comparison diagram before and after the pre-service of different sample pictures, (a) subgraph is former figure among the figure, and (b) subgraph is the grey level histogram of original image; (c) subgraph is the image behind the former figure process smothing filtering, and (d) subgraph is the image after histogram strengthens.Can find out that more clear than original image through the image after strengthening, contrast is stronger.

Fig. 6 to Fig. 9 is respectively the remarkable figure of the said sample picture of Fig. 2 to Fig. 5, is used for the Itti model and seeks the calculating of noting the force.

Figure 10 to Figure 13 is respectively the sample picture comparison diagram as a result that sample picture adopts watershed segmentation algorithm and other partitioning algorithms commonly used to cut apart to Fig. 5 among Fig. 2.Wherein, (a) subgraph is an original image; (b) subgraph is the segmentation effect of FCM (Fuzzy C Means) algorithm, (c) is the segmentation effect of SM_Ath (Segmentation method based onadaptive threshold) algorithm, and (d) subgraph is the segmentation effect of SM_Edge (Segmentation method based on Edge) algorithm; (e) be the segmentation effect of SM_Reg (Segmentation method based on region) algorithm, (f) be the segmentation effect of the inventive method.Can find out that the cut zone profile of this method is more obvious, is easy to extract area-of-interest.

Figure 14 is the area-of-interest mask that sample picture utilization this method is extracted automatically among Fig. 2, and the elementary contour that can find out bear is all by correct extracting, and effect is better.

The area-of-interest mask that Figure 15 extracts for Fig. 3 sample picture utilization this method automatically, the hull partial contour is capped basically, and belt ship top is not found, and effect is not so good.

Figure 16 and Figure 17 are respectively Fig. 4 sample picture and the automatic area-of-interest mask that extracts of Fig. 5 sample picture utilization this method, and people's elementary contour is all come out by correct extraction, and effect is better.

Claims (5)

1. an area-of-interest exacting method is characterized in that, comprises following step:

Step 1: original image is carried out pre-service;

Concrete pre-treatment step is following:

Step 1.1: calculate each pixel in the image (x, pixel value s y) _i, wherein, i is the pixel subscript;

Step 1.2: the number of supposing the pixel in the image is N, calculates the pixel average m of entire image:

$m=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{s}_i---\left(1\right)$

Wherein: total number of pixels of N presentation video, s _iRepresent i gray values of pixel points;

Step 1.3: according to average m and pixel (x, pixel value s y) _i, the departure degree of each pixel and its average during equilibrium is published picture and looked like:

w _i＝‖s _i-m‖ ²(2)

Step 1.4: setup parameter K, as the departure degree w of image pixel value and mean value _iDuring less than K, just on the pixel value of its neighborhood, add K; If departure degree w _iMore than or equal to K, the pixel value of its neighborhood point equals the pixel value of pixel;

Step 2: original image is calculated significantly figure;

At first pretreated image is carried out filtering;

Carry out color space conversion then, with filtered image by the RGB color space conversion to corresponding HIS color space;

Original image is calculated significantly figure, and it is following that the tone overall situation correlative value of calculating pixel x gets formula:

$S_h\left(x\right)=\underset{V_h=V_h\min }{\overset{V_h\max }{\mathrm{\Σ}}}\left(\right|H\left(x\right)-V_h|\·{\mathrm{hist}}_h\left(v_h\right))---\left(4\right)$

$S_s\left(x\right)=\underset{V_s=V_s\min }{\overset{V_s\max }{\mathrm{\Σ}}}\left(\right|\mathrm{Sue}\left(x\right)-V_s|\·{\mathrm{hist}}_s\left(v_s\right))---\left(5\right)$

$S_i\left(x\right)=\underset{V_i=V_i\min }{\overset{V_i\max }{\mathrm{\Σ}}}\left(\right|I\left(x\right)-V_i|\·{\mathrm{hist}}_i\left(v_i\right))---\left(6\right)$

$S_x=\sqrt{S_h{\left(x\right)}^2+S_s{\left(x\right)}^2+S_i{\left(x\right)}^2}---\left(7\right)$

Wherein: in the formula (4), S _h(x) significance of the h component of remarked pixel x, V _hThe minimum value of h component in the min presentation video, V _hThe maximal value of the h component of max remarked pixel x, hist _hThe statistics with histogram result of h component in the presentation video, in the formula (5), S _s(x) significance of the s component of remarked pixel x, V _sThe minimum value of s component in the min presentation video, V _sThe maximal value of the s component of max remarked pixel x, hist _sThe statistics with histogram result of presentation video s component, in the formula (6), S _i(x) significance of the h component of remarked pixel x, V _iThe minimum value of i component in the min presentation video, V _iThe maximal value of the i component of max remarked pixel x, hist _iThe statistics with histogram result of presentation video i component; In the formula (7),, utilize formula (7) to obtain each pixel conspicuousness in the image with the synthesis result significance as a whole of above-mentioned three component significances;

Step 3: original image is cut apart with the watershed segmentation algorithm;

Detailed process is following:

M ₁, M ₂... M _RPresentation video g (x, the coordinate set of local minimum point y), C (M _i) for being positioned at and local minimum M _iThe set of the coordinate of point in the catchment basin that interrelates; Min and max represent g (x, minimum value y) and maximal value; T [n] denotation coordination (s, set t), wherein g (s, t)＜n, that is:

T [n] be g (x, the coordinate set of the point in y), the point in the set all be positioned at plane g (x, y)=below of n;

C _n(M _i) coordinate set of expression catchment basin mid point; C _n(M _i) be the bianry image that provides by following formula:

C _n(M _i)＝C(M _i)∩T[n] (9)

Wherein, C _n(M _i) expression C (M _i) with the common factor of T [n], that is, if (x, y) ∈ C (M _i) and ((x y) has C then in the position for x, y) ∈ T [n] _n(M _i)=1, otherwise, C _n(M _i)=0;

C [n] is illustrated in the intersection of the part that n stage catchment basin do not had by water logging:

$c\left[n\right]=\underset{j=1}{\overset{R}{\∪}}{C}_n\left(M_j\right)---\left(10\right)$

Wherein, R is total number of local minimum; Make C [max+1] be the intersection of all catchment basins then:

$C[\max +1]=\underset{j=1}{\overset{R}{\∪}}C\left(M_j\right)---\left(11\right)$

Wherein, max is g (x, maximal value y), the max pixel value that promptly occurs in the image;

When beginning, the algorithm of looking for watershed line sets C [min+1]=T [min+1]; Algorithm gets into recursive call then, supposes when n goes on foot, and has constructed C [n-1]; The process of trying to achieve C [n] according to C [n-1] is following: make Q represent the set of UNICOM's component among the T [n]; Then, for each component q ∈ Q of UNICOM [n], following three kinds of possibilities are arranged:

(a) q ∩ C [n-1] is empty;

(b) q ∩ C [n-1] comprises UNICOM's component among the C [n-1];

(c) q ∩ C [n-1] comprises unnecessary one the UNICOM's component of C [n-1];

When running into a up-to-date minimum value eligible (a), then incorporate q into C [n-1] and constitute C [n]; When q was arranged in the catchment basin of some local minimum formation, eligible (b) integrated with C [n-1] with q and constitute C [n] this moment; When running into the topographical crest of two or more catchment basins of all or part of separation, eligible (c);

Step 4:, calculate the degree interested in each zone according to remarkable figure and segmentation result;

The size that at first will significantly scheme is shrunk to the size the same with original image in proportion, calculates the degree interested in each zone:

$\mathrm{Interest}\left(R_i\right)={\mathrm{\γ}}_1\×\frac{\mathrm{Num}\left(R_i\right)}{\mathrm{Num}\left(R\right)}+{\mathrm{\γ}}_2\×\frac{\underset{j=1}{\overset{\mathrm{Num}\left(R_i\right)}{\mathrm{\Σ}}}{r}_j}{\mathrm{Num}\left(R_i\right)}---\left(12\right)$

Wherein, γ ₁+ γ ₂=1, represent the weights coefficient of region area and average degree interested, Num (R respectively _i) and Num (R) the pixel number of being with not regional i and overall region to comprise respectively, r _jCertain any degree interested in the expression zone;

Step 5: carry out area-of-interest and detect;

Area-of-interest detects and comprises image segmentation and two parts of interest tolerance:

Region detecting part is divided with gradient watershed transform split image, and (x y) has described the object grey scale change to gradient G, is expressed as

$G(x,y)=\sqrt{{(I(x,y)*G_x)}^2+{(I(x,y)*G_y)}^2}---\left(13\right)$

Wherein: (x y) is gray-scale map, G to I _x, G _yBe the sobel edge mask and At first image segmentation is become several regions, then gradient image is carried out the watershed divide mark and produce mark matrix L rgb, select to note that with focus-of-attention the zone produces mask mask figure at last according to image gradient;

Interest tolerance part adopts vision noticing mechanism to build interest-degree; Choose the watershed segmentation result; Use the Itti attention model to calculate the interest-degree in focus-of-attention tolerance zone, characteristic remarkable property obtains through the difference of gaussian sampling of computed image regional center and periphery, and formula is:

$\mathrm{DOG}(x,y)=\frac{1}{2{\mathrm{\π\σ}}_c^2}\exp |-\frac{x^2+y^2}{2{\mathrm{\σ}}_c^2}|-\frac{1}{2{\mathrm{\π\σ}}_s^2}\exp |-\frac{x^2+y^2}{2{\mathrm{\σ}}_s^2}|---\left(14\right)$

Wherein, (formula (14) is the expression of a double gauss differential mode type on the whole, σ for x, y) the pixel coordinate of presentation video _cAnd σ _sIt is respectively the parameter of model; Respectively to brightness I, color characteristic figure C and direction character figure O ask the conspicuousness of local component with formula (14), and be last, and whole remarkable figure S is the combination of brightness figure I, color characteristic figure C and direction character figure O, and formula is:

S＝w _i×N(I)+w _c×N(C)+w _o×N(O) (15)

In the formula, N (.) is a normalized factor, is about to a component value and constrains in [0,1] interval w _i, w _cAnd w _oBe respectively the characteristic weights of each component, and w _i+ w _c+ w _o=1;

After significantly being schemed S, next step need find the focus-of-attention among the remarkable figure, and the selection of focus-of-attention is to obtain through the neural net method that the victor is a king with shifting, and formula is:

$V(t+\mathrm{\δt})=[1-\frac{\mathrm{\δt}}{\mathrm{CR}}]V\left(t\right)+\frac{\mathrm{\δt}}{C}I\left(t\right)---\left(16\right)$

In the formula, C representes electric capacity, and R is a resistance, and V is a mode voltage, and following formula is represented the output voltage V (t) in the known t moment and the mode voltage V (t+ δ t) that input current I (t) produces behind time δ t, through after a while, and the voltage integrating meter granting of generation; When choosing focus-of-attention to image, with the integration granting neuron array of significantly scheming to regard as a two dimension, the value of pixel among the corresponding significantly figure of neuronic input current; Each neuronic film in will significantly scheming is then led through electricity and is converted neuronic input current in the WTA network to; The WTA network also be also should two dimension integration provide neuron town Liu, it is littler than neuronic time constant among the remarkable figure to play neuronic time constant, it is faster than the neuron among the remarkable figure that current potential rises to; Like this, the neuron in the WTA network always produces earlier than the neuron among the remarkable figure and provides, and the neuron that generation is at first provided is just corresponding to significantly significantly being worth maximum neuron, i.e. focus-of-attention among the figure;

Before focus-of-attention shifts, need choose next focus-of-attention, return inhibition with a width of cloth and scheme to suppress current region, make notice turn to next zone; When carrying out the k time focus-of-attention transfer, return inhibition figure IR at the k width of cloth ^kIn, will belong to interior ground of (k-1) individual marking area pixel value and all be changed to 0, the pixel value of remaining position returns inhibition figure IR with respect to (k-1) width of cloth ^K-1Constant, as shown in the formula:

IR ⁰(x，y)＝1

In the formula, R ^kBe k marking area: (x y) is coordinate points; IR ^k(x, when y) k focus-of-attention chosen in expression, the k width of cloth returned coordinate points among the inhibition figure (x, value y); The area-of-interest of the public definite image of degree regional interested that calculates according to the focus-of-attention that finds and step 4 at last.

2. a kind of area-of-interest exacting method according to claim 1 is characterized in that, parameter K is 1/10 of an image average in the step 1.

3. a kind of area-of-interest exacting method according to claim 1; It is characterized in that median filtering method is adopted in the filtering in the step 2, is specially: earlier a template pixel that contains odd number is roamed in the drawings; And its template center overlapped with certain location of pixels in the image; Read each the respective pixel gray-scale value in the template, then these gray-scale values are sorted from small to large, find out its intermediate value and give the pixel of the center of corresponding templates it.

4. a kind of area-of-interest exacting method according to claim 1 is characterized in that, in the step 2 with filtered image by the RGB color space conversion to corresponding HIS color space, be specially:

$\left\{\begin{array}{c}H=1/360[90-\arctan (F/3)+\{0,G>B;180,G>B\left\}\right]\\ S=1-[\min (R,G,B)/I]\\ I=(R+G+B)/3\\ F=(2R-G-B)/(G-B)\end{array}\right.---\left(3\right)$

Wherein: G representes the value of green component in the rgb space, and B representes the value of rgb space Smalt component, and R representes the value of red component in the rgb space; H representes the value of tone component in the HIS space; I representes the value of strength component in the HIS space, and S representes the value of saturation degree in the HIS space, and F is an intermediate value.

5. a kind of area-of-interest exacting method according to claim 1 is characterized in that, σ in the step 5 _cAnd σ _sValue be respectively 0.5 and 1.5.

Images