CN105528783A - Multi-level compact texture feature extraction method and image texture segmentation method - Google Patents

Abstract

The invention provides a multi-level compact texture feature extraction method and an image texture segmentation method. The multi-level compact texture feature extraction method based on fractal dimension comprises following successive steps: direction and scale information of an original image are extracted to obtain sub-bands of the original image; normalization processing is performed to the sub-bands to obtain a feature image having the same resolution ratio of the original image; fractal features are extracted from the feature image.

Description

To compact at many levels texture characteristic extracting method and Image Texture Segmentation method

Technical field

The present invention relates to texture analysis field, more particularly, the present invention relates to a kind of compact at many levels texture characteristic extracting method and corresponding Image Texture Segmentation method based on fractal dimension.

Background technology

Texture Segmentation is the important foundation link of texture analysis, and it is that image is divided into different parts, and every partial interior has close texture.Texture Segmentation object is the border found out between different texture, and gives a same tag to identical texture.The emphasis of Texture Segmentation can effectively distinguish adjacent different texture, correctly finds the border between adjacent texture region.It comprises texture and represents, determines the basic problems such as the edge of cut zone.

Occurring in nature different types of form material generally has different dimensions; Certain corresponding relation is had between gray scale that is fractal and image represents.Fractal dimension is a kind of powerful weighing imaging surface roughness.Fractal dimension is consistent with the judgement of human eye to imaging surface degree of roughness, and is independent of certain limit intrinsic resolution ratio, independent of visual angle, and the expression amount of the material of stable existence.Therefore fractal dimension is as the tolerance of imaging surface degree of irregularity, is often used as the actual parameter distinguishing different texture.

But existingly utilize single fractal dimension to be not sufficient to describe textural characteristics as the method for textural characteristics, reason is:

(1) fractal method is lacked direction selectivity.Be not suitable for the local grain structure of Description Image.

(2) texture that some visual effects are different but has close fractal dimension.

Summary of the invention

Technical matters to be solved by this invention is for there is above-mentioned defect in prior art, a kind of texture characteristic extracting method that compacts at many levels based on fractal dimension is provided, this multi-level compact texture description method can make up the deficiency of single Cancers Fractional Dimension Feature, extracts the fractal multiple dimensioned Orientation Features effectively representing texture.

In order to realize above-mentioned technical purpose, according to the present invention, providing a kind of texture characteristic extracting method that compacts at many levels based on fractal dimension, comprising the following step performed successively:

First step: the direction of extraction original image and dimensional information are to obtain the subband of original image;

Second step: perform subband normalized to obtain the characteristic image with original image equal resolution;

Third step: fractal characteristic is extracted to characteristic image.

Preferably, in a first step, employing steerable pyramid carries out the filtering on N number of yardstick, a D direction to original image I, obtain N × D subband, each subband { B _n,d(I) } represent, n=0.1...N-1; D=0.1..D-1, wherein n represents yardstick, and d represents direction.

Preferably, in the second step, to the N number of subband { B on same direction _n,d(I), n=0.1...N-1} carries out normalization process, obtains the characteristic image F of D and original image equal resolution altogether _i, i=0,1...D-1.

Preferably, in third step, calculate characteristic image F _ilocal difference box-counting dimension, and adopt Fd _i, i=0,1...D-1 representation feature image F _ieach local difference box-counting dimension, the D dimensional feature vector T obtaining pixel j is:

T(j)＝{Fd ₀(i),...Fd _D-1(i)} ^T。

According to the present invention, additionally provide a kind of Image Texture Segmentation method, comprise the following step performed successively:

First Texture Segmentation step: set up a chart for given image, the weight matrix of design of graphics picture and degree matrix;

Second Texture Segmentation step: utilize the weight matrix built to obtain the second little eigenvector with degree matrix.

Texture segmentation step: utilize the second little eigenvector that figure is divided into two parts.

Preferably, described Image Texture Segmentation method also comprises the 4th Texture Segmentation step: judge whether to need to split image further, if need segmentation image further, again performs the first Texture Segmentation step, the second Texture Segmentation step and texture segmentation step.

Preferably, in the first Texture Segmentation step, the weight matrix W according to following formula design of graphics picture:

The similarity of setting two of image between pixel i, j

aff(i,j)＝exp{-dist(T(i),T(j)) ²/2σ ²}

In formula, dist (T (i), T (j)) is the Euclidean distance between D dimensional vector T (i), T (j)

Weight matrix W:

Wherein r is the radius of neighbourhood of setting, and x is the spatial domain positional information of pixel.

Preferably, in the second Texture Segmentation step, by solving the proper vector of following formula to obtain the second little eigenvector:

Accompanying drawing explanation

By reference to the accompanying drawings, and by reference to detailed description below, will more easily there is more complete understanding to the present invention and more easily understand its adjoint advantage and feature, wherein:

Fig. 1 schematically shows according to the preferred embodiment of the invention based on the Texture Segmentation block diagram of the texture characteristic extracting method that compacts at many levels of fractal dimension.

Fig. 2 schematically shows according to the preferred embodiment of the invention based on the process flow diagram of the texture characteristic extracting method that compacts at many levels of fractal dimension.

Fig. 3 schematically shows according to the preferred embodiment of the invention based on the feature extraction block diagram of the texture characteristic extracting method that compacts at many levels of fractal dimension.

Fig. 4 schematically shows the process flow diagram of Image Texture Segmentation method according to the preferred embodiment of the invention.

It should be noted that, accompanying drawing is for illustration of the present invention, and unrestricted the present invention.Note, represent that the accompanying drawing of structure may not be draw in proportion.Further, in accompanying drawing, identical or similar element indicates identical or similar label.

Embodiment

In order to make content of the present invention clearly with understandable, below in conjunction with specific embodiments and the drawings, content of the present invention is described in detail.

Texture is area attribute, closely related with image resolution ratio.Only just can be perceived under certain yardstick.Psychophysiology experiment shows that the result of visually-perceptible is with different levels institutional framework, i.e. the result of scale selection.Under large scale, the primary structure of objects in images can be caught; Details under small scale in energy perceptual image.The mankind are in the identification mission of texture, and most important three dimensions are directivity, periodicity and randomness, and wherein direction is factor particularly important in texture perception.Therefore, what describe texture must possess multiple dimensioned, multi-direction characteristic; Texture intrinsic characteristic can be weighed; And similar texture has compactness in feature space.Thus, fractal theory combines with direction pyramid theory by the present invention, extracts the image fractal dimension with directional information, effectively represents the multiple dimensioned Orientation Features of texture; Adopt normalized cut as sorter again, as shown in Figure 1.

Fig. 2 schematically shows according to the preferred embodiment of the invention based on the process flow diagram of the texture characteristic extracting method that compacts at many levels of fractal dimension; Fig. 3 schematically shows according to the preferred embodiment of the invention based on the feature extraction block diagram of the texture characteristic extracting method that compacts at many levels of fractal dimension.

As shown in Figures 2 and 3, comprise based on the texture characteristic extracting method that compacts at many levels of fractal dimension according to the preferred embodiment of the invention:

First step S1: the direction of extraction original image and dimensional information are to obtain the subband of original image; Particularly, such as, employing steerable pyramid carries out the filtering on N number of yardstick, a D direction to original image I, obtain N × D subband, each subband { B _n,d(I) } represent, n=0.1...N-1; D=0.1..D-1.Wherein n represents yardstick, and d represents direction.

Second step S2: perform subband normalized to obtain the characteristic image with original image equal resolution; Such as, particularly, to the N number of subband { B on same direction _n,d(I), n=0.1...N-1} carries out normalization process, obtains the characteristic image F of D and original image equal resolution altogether _i, i=0,1...D-1.

Third step S3: fractal characteristic is extracted to characteristic image; Particularly, such as, characteristic image F is calculated _ilocal difference box-counting dimension, and adopt Fd _i, i=0,1...D-1 representation feature image F _ieach local difference box-counting dimension, the D dimensional feature vector T (expression fractal characteristic) obtaining pixel j is:

T(j)＝{Fd ₀(i),...Fd _D-1(i)} ^T(1)

Based on said method, Texture Segmentation process is according to the preferred embodiment of the invention discussed further below.

Adopt normalized cut method as the sorter of Texture Segmentation, be a weighted undirected graph by image mapped, image segmentation problem be converted into the division to weighted undirected graph.If two pixels are similar, the weights on the limit between so corresponding two nodes are just larger, and vice versa, by removing less associated weight limit, Iamge Segmentation is become each subgraph be connected that inner weight is larger, each subgraph just corresponds to the part of a segmentation.

Image can correspond to a weighted undirected graph G=(V, E), and G is original image, and V is all nodes in G, and corresponding to the pixel of image, the connection edge between node represents with E.Connection weights W (i, j) is the similarity function between node i and node j.G=(V, E) can be divided into two mutually disjoint part A, B.By removing A, the fillet between these two parts of B, makes:

The two-part different degree (disassociation) of A, B is:

Wherein cut (A, B) represents A, all connection weights between B and, be defined as follows:

Assoc (A, V) represents the connection weights sum in A between all nodes and V (all nodes).

Make the segmentation that Ncut (A, B) value is minimum, be referred to as normalized cut.Normalization similarity in definition class:

Normalized cut is that two classes split are separated as far as possible, wishes that two class inside are intensive as best one can simultaneously.Namely, while minimizing Ncut (A, B), Nassoc (A, B) is maximized.

For the image I of a width N × N, the size of weight matrix W is N ²× N ².I-th row just represents the weights coefficient between i-th node and other all nodes.Be added by every row element of similarity matrix, namely obtain the degree of this node, diagonal matrix degree of the being matrix formed for diagonal element with all angle value, represents with D.Diagram root is converted into the second little eigenvector problem solving following equation to the implementation procedure of problem of G segmentation.

The weight matrix W of design of graphics picture mainly comprises following three steps:

The first step: with filtering can be carried out to image I by tuning pyramid, obtain the characteristic image { B of filtered one group of different scale, different directions _n,d(I) }.The subband of equidirectional different resolution has self-similarity.But such characteristic image can not ensure that similar node has higher weights and connects.

Second step: calculate characteristic image { B again _n,d(I) the fractal dimension Fd of every bit }.Occurring in nature different types of form material generally has different dimensions, and same fractal material generally has identical dimension in different regions.The fractal dimension of every bit on texture image can be calculated, extract the feature with similarity.The eigenvector of pixel j point as shown in formula (1), T (j)={ Fd ₀(i) ... Fd _d-1(i) } ^t, i=0,1...D-1.

3rd step: the formula (8) of the similarity between pixel i, j defines:

aff(i,j)＝exp{-dist(T(i),T(j)) ²/2σ ²}(8)

In formula, dist (T (i), T (j)) is the Euclidean distance between D dimensional vector T (i), T (j).

Simultaneously in conjunction with the spatial domain positional information x of pixel.Generally speaking, similar to pixel point is usually located in its neighborhood.If the distance of two pixels exceedes the contiguous range preset, think that the weights between them are zero.

Weights similarity matrix W is defined as follows:

Wherein r is the radius of neighbourhood of setting.

In sum, Fig. 4 schematically shows the process flow diagram of Image Texture Segmentation method according to the preferred embodiment of the invention, and as shown in Figure 4, Image Texture Segmentation method key step of the present invention is as follows:

First Texture Segmentation step S10: set up a chart G for given image, according to weight matrix W and the degree matrix D of formula (8), (9) design of graphics picture.

Second Texture Segmentation step S20: utilize the weight matrix and degree matrix that build, by the proper vector of solution formula (7) to obtain the second little eigenvector.

Texture segmentation step S30: utilize the second little eigenvector that figure is divided into two parts.

4th Texture Segmentation step S40: judge whether to need to split image further, if need segmentation image further, recursive call algorithm (i.e. each Texture Segmentation step above-mentioned) is split again, otherwise process terminates.

This shows, texture is that visually-perceptible provides key message, to the field such as pattern-recognition and computer vision important in inhibiting.Textural characteristics represents it is basis and the prerequisite of texture analysis.Need the intrinsic propesties considering texture on the one hand, also will consider parsimony and the completeness of feature on the other hand.The textural characteristics method for expressing based on fractal dimension and anisotropic filter that the present invention proposes can be effectively applied to Texture Segmentation.

In addition, it should be noted that, unless stated otherwise or point out, otherwise the term " first " in instructions, " second ", " the 3rd " etc. describe only for distinguishing each assembly, element, step etc. in instructions, instead of for representing logical relation between each assembly, element, step or ordinal relation etc.

Be understandable that, although the present invention with preferred embodiment disclose as above, but above-described embodiment and be not used to limit the present invention.For any those of ordinary skill in the art, do not departing under technical solution of the present invention ambit, the technology contents of above-mentioned announcement all can be utilized to make many possible variations and modification to technical solution of the present invention, or be revised as the Equivalent embodiments of equivalent variations.Therefore, every content not departing from technical solution of the present invention, according to technical spirit of the present invention to any simple modification made for any of the above embodiments, equivalent variations and modification, all still belongs in the scope of technical solution of the present invention protection.

Claims (8)

1., based on the texture characteristic extracting method that compacts at many levels of fractal dimension, it is characterized in that comprising the following step performed successively:

First step: the direction of extraction original image and dimensional information are to obtain the subband of original image;

Second step: perform subband normalized to obtain the characteristic image with original image equal resolution;

Third step: fractal characteristic is extracted to characteristic image.

2. the texture characteristic extracting method that compacts at many levels based on fractal dimension according to claim 1, it is characterized in that, in a first step, employing steerable pyramid carries out the filtering on N number of yardstick, a D direction to original image I, obtain N × D subband, each subband { B _n,d(I) } represent, n=0.1...N-1; D=0.1..D-1, wherein n represents yardstick, and d represents direction.

3. the texture characteristic extracting method that compacts at many levels based on fractal dimension according to claim 1 and 2, is characterized in that, in the second step, to the N number of subband { B on same direction _n,d(I), n=0.1...N-1} carries out normalization process, obtains the characteristic image F of D and original image equal resolution altogether _i, i=0,1...D-1.

4. the texture characteristic extracting method that compacts at many levels based on fractal dimension according to claim 1 and 2, is characterized in that, in third step, calculates characteristic image F _ilocal difference box-counting dimension, and adopt Fd _i, i=0,1...D-1 representation feature image F _ieach local difference box-counting dimension, the D dimensional feature vector T obtaining pixel j is:

T(j)＝{Fd ₀(i),…Fd _D-1(i)} ^T。

5. an Image Texture Segmentation method, is characterized in that comprising the following step performed successively:

First Texture Segmentation step: set up a chart for given image, the weight matrix of design of graphics picture and degree matrix;

Second Texture Segmentation step: utilize the weight matrix built to obtain the second little eigenvector with degree matrix.

Texture segmentation step: utilize the second little eigenvector that figure is divided into two parts.

6. Image Texture Segmentation method according to claim 5, characterized by further comprising the 4th Texture Segmentation step: judge whether to need to split image further, if need segmentation image further, again perform the first Texture Segmentation step, the second Texture Segmentation step and texture segmentation step.

7. the Image Texture Segmentation method according to claim 5 or 6, is characterized in that, in the first Texture Segmentation step, and the weight matrix according to following formula design of graphics picture:

The similarity of setting two of image between pixel i, j

aff(i,j)＝exp{-dist(T(i),T(j)) ²/2σ ²}

In formula, dist (T (i), T (j)) is the Euclidean distance between D dimensional vector T (i), T (j)

$dist\left(T\right(i),T(j\left)\right)=\sqrt{\underset{l=0}{\overset{D-1}{\Σ}}{\left({\mathrm{Fd}}_l\right(i)-{\mathrm{Fd}}_l(j\left)\right)}^2}$

Weight matrix:

$W_{ij}=aff(i,j)\×\left\{\begin{array}{cc}|x\left(i\right)-x\left(j\right)|,& |x\left(i\right)-x\left(j\right)|\≤r\\ 0,& |x\left(i\right)-x\left(j\right)|>r\end{array}\right\}$

Wherein r is the radius of neighbourhood of setting, and x is the spatial domain positional information of pixel.

8. the Image Texture Segmentation method according to claim 5 or 6, is characterized in that, in the second Texture Segmentation step, by solving the proper vector of following formula to obtain the second little eigenvector:

$D^{-\frac{1}{2}}(D-W)D^{-\frac{1}{2}}Z=\mathrm{\λ}Z.$