CN105741294A

CN105741294A - Object-quantity-controllable image segmentation method

Info

Publication number: CN105741294A
Application number: CN201610069837.0A
Authority: CN
Inventors: 施文灶
Original assignee: Fujian Normal University
Current assignee: Fujian Normal University
Priority date: 2016-02-01
Filing date: 2016-02-01
Publication date: 2016-07-06

Abstract

The invention relates to an object-quantity-controllable image segmentation method, which comprises the following steps of 1, constructing an image I to be processed into a graph G; 2, for a segmentation objective with the preset object quantity being k, converting k objects segmented from the image I into k feature vectors of a solving graph G; 3, merging and constructing the k feature vectors solved in the step 2 into an MN*k matrix Mtr; 4, taking vectors according to the lines of the matrix Mtr in the step 3 to obtain MN 1*k line vectors and feature vectors corresponding to each pixel in the image I; 5, clustering the feature vector of each pixel in the step 4 by a clustering algorithm to obtain k segmentation objects; and 6, performing visual expression on k clustering objects in the step 5. The method solves the problem of image division object uncontrollability, and has high segmentation accuracy. The over-segmentation and the under-segmentation of the image can be well realized, and the method is applied to fixed-quantity objective extraction.

Description

The image partition method that a kind of number of objects is controlled

Technical field

The present invention relates to a kind of digital image processing field, the image partition method that specifically a kind of number of objects is controlled.

Background technology

Image segmentation is a classic problem in image procossing, is also the basic fundamental in image procossing and computer vision field.The research of image segmentation is constantly subjected to the great attention of people for many years, and partitioning algorithm also emerges in an endless stream, for the classification foundation also disunity of image segmentation algorithm.The selection of image partition method, is largely dependent upon the variable factor in specific image, imaging mode and imaging and invariant factor (such as noise and texture etc.), and these all can affect follow-up segmentation to a great extent.It addition, current partitioning algorithm there is also one problem to be solved, namely the quantity of cutting object is uncontrollable, it is difficult to adapt to some application scenario.

Summary of the invention

The invention provides the image partition method that a kind of number of objects is controlled, utilize Graph-theoretical Approach, the problem of dividing the image into is converted into characteristic vector Solve problems, has the advantage that cutting object is controlled and segmentation accuracy is high.

Target by realizing the present invention be the technical scheme is that method comprises the following steps:

Step 1: the pending image I tectonic ore-forming G of M × N will be of a size of；

Step 2: be the segmentation object of k for default number of objects, is converted into, by I is divided into k object, k the characteristic vector solving figure G；

Step 3: k the characteristic vector (column vector of MN × 1) tried to achieve in step 2 is merged the matrix Mtr being configured to MN × k；

Step 4: by the row amount of orientation of the matrix Mtr in step 3, obtain the row vector of MN 1 × k, obtains each pixel characteristic of correspondence vector in image I；

Step 5: utilize clustering algorithm that the characteristic vector of each pixel in step 4 is carried out cluster and obtain k cutting object；

Step 6: k clustering object in step 5 is carried out visual representation.

The building method of the figure G in described step 1 is: using the summit as figure of each pixel in pending image I, if the distance between two pixels is less than threshold value d, be then attached with a nonoriented edge, the weight w on limit_M(m, n) is arranged by below equation:

w_{M} (m, n) = \sqrt{w_{D} (m, n) \times w_{I} (m, n)}

Wherein, w_D(m, n)=exp (-| | L_m-L_n||²), represent the proximity relations of two pixels；w_I(m, n)=exp (-(I_m-I_n)²), represent the close degree of value of two pixels.

The method of k the characteristic vector solving figure G in described step 2 utilizes below equation:

(D-W) y=λ Dy

Wherein D is diagonal matrix,

W = [\begin{matrix} w (1, 1) & ... & w (1, M N) \\ ... & ... & ... \\ w (M N, 1) & ... & w (M N, M N) \end{matrix}],

Y is characteristic vector, and λ is eigenvalue.

Clustering algorithm in described step 5 is k mean algorithm.

The invention has the beneficial effects as follows: solve the uncontrollable problem of image cutting object, there is higher segmentation accuracy.Over-segmentation and the less divided of image can be realized well, be applied to the Objective extraction of fixed qty.

Accompanying drawing explanation

Fig. 1 is the overall process flow figure of the present invention.

Detailed description of the invention

The specific embodiment of the present invention is described in detail below in conjunction with accompanying drawing.

In step 101, input is of a size of the pending image I of M × N.

In step 102, by pending image I tectonic ore-forming G, by the summit as figure of each pixel in image I, if the distance between two pixels is less than threshold value 20, then it is attached with a nonoriented edge, the weight w on limit_M(m, n) is arranged by below equation:

w_{M} (m, n) = \sqrt{w_{D} (m, n) \times w_{I} (m, n)}

In step 103, being the segmentation object of k for default number of objects, the method for k the characteristic vector solving figure G utilizes below equation:

(D-W) y=λ Dy

Wherein D is diagonal matrix,

W = [\begin{matrix} w (1, 1) & ... & w (1, M N) \\ ... & ... & ... \\ w (M N, 1) & ... & w (M N, M N) \end{matrix}],

Y is characteristic vector, and λ is eigenvalue.

In step 104, k the characteristic vector (column vector of MN × 1) tried to achieve in step 103 is merged the matrix Mtr being configured to MN × k, and the row amount of orientation to matrix Mtr, obtain the row vector of MN 1 × k, obtain each pixel characteristic of correspondence vector in image I.

In step 105, utilize k means clustering algorithm that the characteristic vector of each pixel in step 104 is carried out cluster and obtain k cutting object.

In step 106, k clustering object in step 105 is carried out visual representation.

Claims

1. the image partition method that a number of objects is controlled, it is characterised in that comprise the following steps:

Step 6: k clustering object in step 5 is carried out visual representation.

2. the image partition method that a kind of number of objects according to claim 1 is controlled, it is characterized in that the building method of the figure G in step 1 is: using the summit as figure of each pixel in pending image I, if the distance between two pixels is less than threshold value d, then it is attached with a nonoriented edge, the weight w on limit_M(m, n) is arranged by below equation:

w_{M} (m, n) = \sqrt{w_{D} (m, n) \times w_{I} (m, n)}

3. the image partition method that a kind of number of objects according to claim 1 is controlled, it is characterised in that the method for k the characteristic vector solving figure G in step 2 utilizes below equation:

(D-W) y=λ Dy

Wherein D is diagonal matrix,

W = [\begin{matrix} w (1, 1) & ... & w (1, M N) \\ ... & ... & ... \\ w (M N, 1) & ... & w (M N, M N) \end{matrix}],

Y is characteristic vector, and λ is eigenvalue.

4. the image partition method that a kind of number of objects according to claim 1 is controlled, it is characterised in that the clustering algorithm in step 5 is k mean algorithm.