CN101059870A - Image cutting method based on attribute histogram - Google Patents

Abstract

An image segmentation method based on character histogram belongs to digit image processing method for resolving the problem of prior method to improve segment accuracy and real-time processing speed. The invention can be used in target check and recognition, comprising (1) image input, (2) image compression, (3) statistic process that calculating the gray spatial distribution density probability matrix relative to the compressed image, (4) histogram extraction for extracting one-dimension gray spatial distribution character histogram relative to the compressed image, (5) threshold value search for using KSW maximum entropy image segmentation method to find the segment threshold value of image, (6) image segmentation for segmenting the compressed image based on the set threshold value. The invention applies character histogram into KSW maximum entropy image segmentation method, uses the gray spatial distribution character as the effective character of character histogram, to effectively segment target with confirmed real-time property.

Description

A kind of image partition method based on attribute histogram

Technical field

The invention belongs to digital image processing method, be specifically related to a kind of image partition method, be used for target detection and identification based on attribute histogram.

Background technology

Image segmentation is that piece image is resolved into some not set of overlapping region mutually, is content very important in the image processing field, also is the important component part of automatic target identification.The correctness of image segmentation and adaptivity affect the intelligent degree of target detection and identification to a certain extent, and the processing speed of image segmentation algorithm has also influenced the real-time of its application.

In numerous image partition methods, threshold method utilizes target and the difference of background on gamma characteristic that will extract in the image, come image is cut apart by choosing suitable gray threshold, thereby target is distinguished from background, because of its realization simply (does not need input parameter, do not need supervision), calculated amount is little, steady performance is widely adopted.Entropy is the sign of average information, and the early 1980s people begin to utilize in the information theory notion of entropy to choose segmentation threshold.T.Pun.A new method for grey-level picture thresholding using the entropyof the histogram.Signal Processing in 1980,1980, propose maximum a posteriori entropy upper bound method at 2 (3): 223～237, nineteen eighty-two G.Johannsen, J.Bille.A threshold selection method usinginformation measures.Proceedings of the 6th International Conference onPattern Recognition, Munich, Germany, 1982, the minimal information cross-correlation method that 1:140～142 propose has also adopted the entropy of image gray levels, J.N.Kapue in 1985, P.K.Sahoo, A.K.C.Wong.A new method for gray-level picture thresholding using the entropy ofthe histogram.Computer Vision, Graphics, and Image Processing, 1985, proposed maximum entropy image partition method at 29 (3): 273～285, is called for short the KSW method.N.R.Pal in 1989, S.K.Pal.Entropic thresholding.Signal Processing, 1989, introduce the q rank local entropy of image and the notion of conditional entropy at 16 (2): 97～108, maximum second order local entropy and maximal condition entropy selection of threshold method have been proposed, A.S.Abutaleb.Automatic thresholding of gray-levelpictures using two-dimensional entropy.Computer Vision in 1989, Graphics andImage Processing, introduced a kind of maximum entropy threshold choosing method of considering homogeneity and shape at 1989,47 (1): 22～32.The KSW method is effectively the simplest, most widely used in these entropy methods, is not desirable double-peak shape but work as image grey level histogram, and then the segmentation effect of KSW method can be affected.

Traditional one dimension image grey level histogram has only been added up the probability of occurrence of gray level in image, do not consider the space distribution information of gray scale, so the existence at noise and edge has greatly influenced segmentation effect.The two dimensional gray histogram had both been added up the intensity profile information of pixel, also utilized the spatial characteristics of pixel, the scope of considering is transformed into the combination of the gray average in the gray scale of this point and its field by the gray-scale value of single-point, A.S.Abutaleb.Automatic thresholdingof gray-level pictures using two-dimensional entropy.Computer Vision, Graphics, and Image Processing, 1989, proposed based on two dimensional gray histogrammic Threshold Segmentation Algorithm at 47 (1): 22～32, with traditional partitioning algorithm ratio, segmentation effect improves.The defective of two dimensional gray histogram partitioning algorithm is a length consuming time, and real-time is poor.

Lu met the spring in 2002, Zhang Dianlun, Guo Haitao. based on the image partition method of attribute histogram and in the sonar image Application in Segmentation. Harbin Engineering University's journal, 2002,23 (3): 1～3 has proposed the notion of attribute histogram, attribute histogram has only considered to have in the image pixel of certain attribute, it is a kind of histogram of incompleteness in a sense, exactly because but this incompleteness, make attribute histogram can give up that some have the interference component or the unconcerned composition of particular problem of certain priori in the histogram of ordinary meaning, and the simplification histogram obtains simplifying and becoming feasible thereby make based on histogrammic image processing method.But people such as Lu Fengchun just is applied to attribute histogram during sonar image cuts apart, and concrete employed attribute lacks general.

Summary of the invention

The invention provides a kind of image partition method based on attribute histogram, purpose is to overcome existing problem in the existing method, under the prerequisite that guarantees real-time, improves the segmentation precision of target.

A kind of image partition method of the present invention based on attribute histogram, order comprises:

(1) image input step, the input original image;

(2) compression of images step, with the number of greyscale levels boil down to of original image its 1/2～1/4, make the intensity profile of target and background area pixels more even;

(3) statistic procedure, statistics compressed images f (x, y) corresponding gray scale space distribution density probability matrix:

Image f (x, y) size is M * N, and number of greyscale levels is m, and its corresponding gray scale space distribution density probability matrix T is that size is m * K ²Two-dimensional matrix, the element T in the matrix (i, j) be illustrated in image f (x, y) in gray level i in image in all K * K neighborhood distribution density be the probability of j, promptly gray level i j time probability occurs in all K * K neighborhood in image, $\underset{i=0}{\overset{m-1}{\mathrm{\Σ}}}\underset{j=1}{\overset{K^2}{\mathrm{\Σ}}}T(i,j)=1;$ I=0,1 ... m-1, j=1,2 ... K ²K=3 or 4;

(4) extract the histogram step, based on gray space distribution density probability matrix T, extract image f (x, y) pairing one dimension gray space properties of distributions histogram, one timing of j value, the one dimension gray space properties of distributions histogram of being constructed by the corresponding column vector of each dimension among the T is defined as:

$p_Q\left(l\right)=\frac{T(l,j)}{\underset{i=0}{\overset{m-1}{\mathrm{\Σ}}}T(i,j)},$

L=0 in the formula, 1 ... each gray level of pixel in the m presentation video; J is a round values between 1～3, represents a certain gray space distribution density value;

(5) determine threshold step,, utilize KSW maximum entropy image partition method to determine image f (x, segmentation threshold y) based on one dimension gray space properties of distributions histogram;

(6) image segmentation step, based on determined threshold value, (x y) is cut apart, and obtains binary image to image f.

Described a kind of image partition method based on attribute histogram is characterized in that: in the described compression of images step, and image pixel original gray level l, the gray level after its compression is l _p, l _pWith the transformational relation of l be:

$l_p=\mathrm{int}\left(\frac{l\×m}{256}\right)\×\frac{256}{m},$

Int () expression rounds operation in the formula, m=128 or 64, the number of greyscale levels that will be compressed into for image.

The intensity profile information that existing two dimensional gray histogram had both been added up pixel, also utilized the spatial characteristics of pixel, the scope of considering is transformed into the gray scale of this point and the gray average combination in its field by the gray-scale value of single-point, with traditional partitioning algorithm ratio, the histogrammic Threshold Segmentation Algorithm of two dimensional gray improves on segmentation effect.

Attribute histogram has only considered to have in the image pixel of certain attribute, it is a kind of histogram of incompleteness in a sense, exactly because but this incompleteness, make attribute histogram can give up that some have the interference component or the unconcerned composition of particular problem of certain priori in the histogram of ordinary meaning, and the simplification histogram, its leaching process is:

(1) be L, size with gray level for all pixels of the image of M * N be considered as domain X={ (x, y) | 0≤x≤M-1,0≤y≤N-1}, wherein (x, y) remarked pixel.If Q is certain property set on the domain X, (x, y) (x y) has certain attribute to ∈ Q remarked pixel.For image X, the attribute histogram on the attribute Q is defined as by P _Q(l)-and discrete figure line that l forms, P _Q(l) be:

$p_Q\left(l\right)=\frac{n_Q\left(l\right)}{N_Q};$

In the formula: n _Q(l) having attribute Q and gray-scale value among the presentation video X is the number of pixels of l, N _QHave the sum of all pixels of attribute Q in the presentation video, have $\underset{l=L_1}{\overset{L_2}{\mathrm{\Σ}}}{p}_Q\left(l\right)=1;$ L wherein ₁, L ₂Be respectively the minimum gradation value and the maximum gradation value of pixel among the property set Q, common 0≤L ₁, L ₂≤ L-1.

(2) key of determining attribute histogram is that the ordinary construction formula of Q is according to particular problem and priori structure attribute collection Q: Q={ (x, y) | Φ [(x, y)] };

In the formula: Φ [(x, y)] (x y) has certain priori Φ to remarked pixel, that is to say to have certain constraint Φ.

The present invention uses for reference the thought of two-dimensional histogram, utilize defined gray space distribution density probability matrix, with the spatial characteristics of pixel effective attribute, proposed effectively to strengthen the one dimension gray space properties of distributions histogram of histogram double-hump characteristics as attribute histogram.

Because target and background area interior pixels are more even, so has more greatly between its respective pixel and the neighborhood territory pixel and may have identical gray-scale value, so the zone of the distribution density probable value of target and background institute corresponding grey scale level j＞1 in defined gray space distribution density probability matrix T has stronger distribution.And the gray level and the space distribution thereof of noise and marginal point pixel are more mixed and disorderly, so the areal distribution of distribution density probable value j＞1 in T of institute's corresponding grey scale level can be more weak.From the angle of one dimension, when one timing of j value, the property set Q of attribute histogram is defined as:

Q={ (x, y) | in all K of image * K neighborhood, have same grey level with j-1 neighborhood territory pixel

Pixel

The one dimension gray space properties of distributions histogram of being constructed by the corresponding column vector of each dimension among the T is defined as:

$p_Q\left(l\right)=\frac{T(l,j)}{\underset{i=0}{\overset{m-1}{\mathrm{\Σ}}}T(i,j)},$

L=0 in the formula, 1 ... each gray level of pixel in the m presentation video; J is an a certain round values between 1 to 3, represents a certain gray space distribution density value;

By the argumentation of front as can be known, in the one dimension gray space properties of distributions histogram of corresponding j＞1, noise and marginal point corresponding gray scale level can be suppressed, and target and background corresponding gray scale level can obtain keeping, thereby histogrammic double-hump characteristics can be strengthened.

Table 1 is based on the comparison consuming time of different partitioning algorithm aircrafts, cell and integrated circuit (IC) image segmentation.

Table 1

With common histogram KSW maximum entropy image partition method, PShi entropy partitioning algorithm is compared with JM entropy partitioning algorithm, carrying out image threshold segmentation method proposed by the invention can effectively improve the effect of cutting apart, and gets rid of the noise spot that is divided into target by mistake, and is about the same on the real-time; Compare with two-dimensional histogram KSW maximum entropy image partition method, the segmentation effect of carrying out image threshold segmentation method proposed by the invention is about the same with it, but tool has great advantage on real-time.

Description of drawings

Fig. 1 is a schematic flow sheet of the present invention;

Fig. 2 is the original image of a width of cloth aircraft;

Fig. 3 is the gray space distribution density probability matrix schematic three dimensional views of aircraft compressed image correspondence;

Fig. 4 is the common histogram of aircraft original image;

Fig. 5 is the one dimension gray space properties of distributions histogram of aircraft compressed image;

Fig. 6 is that the aircraft original image is based on common histogrammic KSW maximum entropy image segmentation result;

Fig. 7 be the aircraft original image based on common histogrammic PShi entropy diagram as segmentation result;

Fig. 8 be the aircraft original image based on common histogrammic JM entropy diagram as segmentation result;

Fig. 9 is the KSW maximum entropy image segmentation result of aircraft original image based on two-dimensional histogram;

Figure 10 is that the aircraft compressed image is based on the histogrammic KSW maximum entropy of one dimension gray space properties of distributions image segmentation result;

Figure 11 is a width of cloth cell original image;

Figure 12 is that the cell original image is based on common histogrammic KSW maximum entropy image segmentation result;

Figure 13 be the cell original image based on common histogrammic PShi entropy diagram as segmentation result;

Figure 14 be the cell original image based on common histogrammic JM entropy diagram as segmentation result;

Figure 15 is the KSW maximum entropy image segmentation result of cell original image based on two-dimensional histogram;

Figure 16 is that the cell compressed image is based on the histogrammic KSW maximum entropy of one dimension gray space properties of distributions image segmentation result;

Figure 17 is width of cloth integrated circuit (IC) original image;

Figure 18 is the KSW maximum entropy image segmentation result of IC original image based on two-dimensional histogram;

Figure 19 be the IC original image based on common histogrammic PShi entropy diagram as segmentation result;

Figure 20 be the IC original image based on common histogrammic JMShi entropy diagram as segmentation result;

Figure 21 is the KSW maximum entropy image segmentation result of IC original image based on two-dimensional histogram;

Figure 22 is that the IC compressed image is based on the histogrammic KSW maximum entropy of one dimension gray space properties of distributions image segmentation result.

Embodiment

The present invention utilizes self-defining one dimension gray space properties of distributions histogram, determines the image segmentation threshold value in conjunction with KSW maximum entropy image partition method, treatment scheme as shown in Figure 1:

(1) image input step, the input original image;

(2) compression of images step is compressed the gray level of original image, makes the intensity profile of target and background area pixels more even; Concrete grammar is as follows:

For the original gray level l in the image, the gray level after its compression is l _p, l _pAs follows with the transformational relation of l:

$l_p=\mathrm{int}\left(\frac{l\×m}{256}\right)\×\frac{256}{m};$

Int () expression rounds operation in the formula, and m=128 or 64 is the number of greyscale levels that image will be compressed into, desirable m=128 in concrete the enforcement, and the grayscale dynamic range that is about to image is by 256 boil down tos 128.

(3) statistics compressed image corresponding gray scale space distribution density probability matrix, this matrix has reflected the space distribution information of gray scale in the image.For size is M * N, and number of greyscale levels is that (x, y), its corresponding gray scale space distribution density probability matrix T is that size is for m * K for the gray level image f of m ²Two-dimensional matrix, the element T in the matrix (i, j) (i=0,1 ... m-1, j=1,2 ... K ²) expression be image f (x, y) in, gray level i in image in all K * K neighborhood distribution density be the probability of j, promptly gray level i j time probability occurs in all K * K field in image.Have for the unit among the T $\underset{i=0}{\overset{m-1}{\mathrm{\Σ}}}\underset{j=1}{\overset{K^2}{\mathrm{\Σ}}}T(i,j)=1.$ Desirable K=3 in concrete the enforcement, promptly the neighborhood size is 3 * 3.Shown in Figure 2 is a width of cloth aircraft image, and shown in Figure 3 is this image corresponding gray scale space distribution density probability matrix.The grey level distribution of sky background is in the scope of [165,175] in the aircraft image of Fig. 2, and T can very clearly observe by intensity profile density probability matrix, and there is very strong distribution in the zone of background corresponding gray scale level k＞1 in T.

(4), extract the pairing one dimension gray space of compressed image properties of distributions histogram based on gray space distribution density probability matrix.From the angle of one dimension, when j value one regularly, the corresponding column vector of each dimension can both structuring one-dimensional gray space properties of distributions histogram among the T, the property set Q that is about to attribute histogram is defined as:

Q={ (x, y) | in all K of image * K neighborhood, have same grey level with j-1 neighborhood territory pixel

Pixel

The one dimension gray space properties of distributions histogram of being constructed by each dimensional vector is defined as:

$p_Q\left(l\right)=\frac{T(l,j)}{\underset{i=0}{\overset{m-1}{\mathrm{\Σ}}}T(i,j)};$

L=0 in the formula, 1 ... each gray level in the m presentation video; J is 1 to K ²Between a certain round values, represent a certain gray space distribution density value.By the argumentation of front as can be known, in the one dimension gray space properties of distributions histogram of corresponding j＞1, noise and marginal point corresponding gray scale level can be suppressed, and target and background corresponding gray scale level can obtain keeping, thereby histogrammic double-hump characteristics can be strengthened.Desirable j=2 in concrete the enforcement, the property set Q that is about to attribute histogram is defined as: Q={ (x, y) | in all 3 * 3 neighborhoods of image, the pixel that has same grey level with 1 neighborhood territory pixel }, shown in Figure 4 is the common histogram of aircraft original image, the pairing one dimension gray space of aircraft compressed image properties of distributions histogram shown in Figure 5.Can see that by the contrast of Fig. 4 and Fig. 5 the defined one dimension gray space of the present invention properties of distributions histogram can effectively strengthen histogrammic double-hump characteristics.

(5), utilize KSW maximum entropy image segmentation algorithm to determine the segmentation threshold OT of image based on one dimension gray space properties of distributions histogram.

Based on the pairing one dimension gray space of aircraft compressed image properties of distributions histogram shown in Figure 5, determine the segmentation threshold OT=130 of aircraft compressed image.

(6) based on determined threshold value, compressed image is cut apart, obtained binary image.

If f (x y) is original image, g (x y) is the binary image after cutting apart, then:

$g(x,y)=\left\{\begin{array}{c}0,f(x,y)<=\mathrm{OT}\\ 255,f(x,y)>\mathrm{OT}\end{array}\right.;$

Fig. 6 is that the aircraft original image is based on common histogrammic KSW maximum entropy image segmentation result.Fig. 7 be the aircraft original image based on common histogrammic PShi entropy diagram as segmentation result.Fig. 8 be the aircraft original image based on common histogrammic JM entropy diagram as segmentation result.Fig. 9 is the KSW maximum entropy image segmentation result of aircraft original image based on two-dimensional histogram.Figure 10 is that the aircraft compressed image is based on the histogrammic KSW maximum entropy of one dimension gray space properties of distributions image segmentation result.

According to above steps, equally can the pair cell image and the integrated circuit (IC) image cut apart.

Figure 11 is a width of cloth cell original image.Figure 12 is that the cell original image is based on common histogrammic KSW maximum entropy image segmentation result.Figure 13 be the cell original image based on common histogrammic PShi entropy diagram as segmentation result.Figure 14 be the cell original image based on common histogrammic JM entropy diagram as segmentation result.Figure 15 is the KSW maximum entropy image segmentation result of cell original image based on two-dimensional histogram.Figure 16 is that the cell compressed image is based on the histogrammic KSW maximum entropy of one dimension gray space properties of distributions image segmentation result.

Figure 17 is a width of cloth IC original image.Figure 18 is the KSW maximum entropy image segmentation result of IC original image based on two-dimensional histogram.Figure 19 be the IC original image based on common histogrammic PShi entropy diagram as segmentation result.Figure 20 be the IC original image based on common histogrammic JMShi entropy diagram as segmentation result.Figure 21 is the KSW maximum entropy image segmentation result of IC original image based on two-dimensional histogram.Figure 22 is that the IC compressed image is based on the histogrammic KSW maximum entropy of one dimension gray space properties of distributions image segmentation result.

Claims (2)

1. image partition method based on attribute histogram, order comprises:

(1) image input step, the input original image;

(2) compression of images step, with the number of greyscale levels boil down to of original image its 1/2～1/4, make the intensity profile of target and background area pixels more even;

(3) statistic procedure, statistics compressed images f (x, y) corresponding gray scale space distribution density probability matrix:

Image f (x, y) size is M * N, and number of greyscale levels is m, and its corresponding gray scale space distribution density probability matrix T is that size is m * K ²Two-dimensional matrix, the element T in the matrix (i, j) be illustrated in image f (x, y) in gray level i in image in all K * K neighborhood distribution density be the probability of j, promptly gray level i j time probability occurs in all K * K neighborhood in image, $\underset{i=0}{\overset{m-1}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{K^2}{\mathrm{\&Sigma;}}}T(i,j)=1;$ I=0,1 ... m-1, j=1,2 ... K ²K=3 or 4;

(4) extract the histogram step, based on gray space distribution density probability matrix T, extract image f (x, y) pairing one dimension gray space properties of distributions histogram, one timing of j value, the one dimension gray space properties of distributions histogram of being constructed by the corresponding column vector of each dimension among the T is defined as:

$p_Q\left(l\right)=\frac{T(l,j)}{\underset{i=0}{\overset{m-1}{\mathrm{\&Sigma;}}}T(i,j)},$

L=0 in the formula, 1 ... each gray level of pixel in the m presentation video; J is a round values between 1～3, represents a certain gray space distribution density value;

(5) determine threshold step,, utilize KSW maximum entropy image partition method to determine image f (x, segmentation threshold y) based on one dimension gray space properties of distributions histogram;

(6) image segmentation step, based on determined threshold value, (x y) is cut apart, and obtains binary image to image f.

2. a kind of image partition method based on attribute histogram as claimed in claim 1 is characterized in that: in the described compression of images step, and image pixel original gray level l, the gray level after its compression is l _p, l _pWith the transformational relation of l be:

$l_p=\mathrm{int}\left(\frac{l\&times;m}{256}\right)\&times;\frac{256}{m},$

Int () expression rounds operation in the formula, m=128 or 64, the number of greyscale levels that will be compressed into for image.

Images