CN103578123B

CN103578123B - A kind of image-region merges method

Info

Publication number: CN103578123B
Application number: CN201310469596.5A
Authority: CN
Inventors: 郑丽颖; 田凯; 石大明
Original assignee: Harbin Engineering University
Current assignee: Harbin Engineering University
Priority date: 2013-10-10
Filing date: 2013-10-10
Publication date: 2016-06-29
Anticipated expiration: 2033-10-10
Also published as: CN103578123A

Abstract

The present invention relates to a kind of image-region and merge method, it is characterised in that: the color histogram map distance of image-region A and image-region B obtains by the following method,Wherein,In formula, p (k) is the color histogram of image-region A, and q (k) is the color histogram of image-region B, wherein k=0,1 ..., N_color-1, N_colorThe number in the color quantizing region that expression color histogram adopts, [A₀,A₁] for the supporting zone of p (k), [B₀,B₁] for the supporting zone of q (k), (p q) represents the distance between two image-region color histograms to D；According to color histogram map distance D (p, size q), it may be judged whether image-region A and image-region B is merged.

Description

Image area merging method

Technical Field

The invention relates to an image region merging method.

Background

Image segmentation plays an important role in pattern recognition and image processing. The results of image segmentation directly determine the performance of object tracking, image understanding, and image compression techniques that encounter objects. Currently, many image processing methods have been proposed, such as: threshold segmentation, feature space clustering techniques, boundary-based segmentation methods, region-based segmentation methods, graph segmentation methods, and the like. In recent years, the graph division method has been receiving more and more attention from researchers. The graph partitioning method mainly studies the node clustering problem, and the mainstream methods include normalized Cut (normalized Cut, N-Cut) and ratio Cut (R-Cut). However, the graph partitioning method has two disadvantages: the required storage space is large and the calculation speed is slow. To overcome these two drawbacks, researchers have proposed a multi-level image segmentation method that combines region-based image segmentation with image segmentation. For example: yang et al propose an image segmentation method based on watershed square and graph theory; tao et al propose an image segmentation method based on Mean Shift (MS) and N-Cut.

Although these region-based image segmentation methods and multi-level image segmentation methods of image segmentation have great advantages in overcoming memory space and computational speed, they generally use simple features (e.g., color mean) to characterize image regions. In multi-level image segmentation, the image area obtained by primary segmentation of an image is usually large, and the features of the image area cannot be completely represented by simple feature vectors. And the research shows that: the color histogram can represent both global and local features of the image color distribution. Thus, the color histogram of the image may show more features. However, the euclidean distance, the histogram intersection distance, or the Bhattacharyya distance, which is suitable for pattern matching and classification, is not suitable for image region merging.

Disclosure of Invention

The invention aims to provide an image area merging method which can effectively realize merging of image areas.

The technical scheme for realizing the invention is as follows:

an image region merging method, characterized by:

the color histogram distance of the image area a and the image area B is obtained by the following method,

wherein,

where p (k) is a color histogram of the image area a, and q (k) is a color histogram of the image area B, where k is 0,1, …, N_color-1，N_colorIndicates the number of color quantization regions used for the color histogram, [ A ]₀,A₁]A support region of p (k) [ B ]₀,B₁]Support region of q (k), D (p, q) represents the distance between the two image region color histograms;

and judging whether the image area A and the image area B are combined or not according to the size of the distance D (p, q) of the color histogram.

Image area a and image area B may be merged when the color histogram distance D (p, q) < a certain threshold.

The invention has the following beneficial effects:

the color histogram distance calculation method provided by the invention fully considers the characteristics of image areas in multi-level image segmentation, has high accuracy of image area combination, and is simple and easy to implement.

Drawings

FIG. 1 is an example image;

fig. 2 is a color histogram of the A, B, C image region of fig. 1.

Detailed Description

Let A, B denote two image regions, p (k) be the color histogram of region a, and q (k) be the color histogram of region B, where k is 0,1, …, N_color-1，N_colorIndicates the number of color quantization regions used for the color histogram, [ A ]₀,A₁]A support region of p (k), i.e., p (k) ≠ 0 if k ∈ [ A ]₀,A₁]，[B₀,B₁]A support region of q (k), i.e., q (k) ≠ 0 if k ∈ [ B [)₀,B₁]. The histogram distance defining area A, B is:

wherein,

in formula (1), D (p, q) represents the distance between histograms p and q, and # represents the intersection of the sets.

Equation (1) indicates that: 1) if region a contains all colors that appear in region B, then the histogram distance between A, B is 0; 2) similarly, if region B contains all the colors that appear in region a, then the histogram distance between A, B is also 0; 3) if region a does not contain any color in region B, then the histogram distance between A, B is also 1; 4) if not the case of 1), 2), 3), the histogram distance between A, B isWherein K₀、K₁Is related to the intersection of the two sets and is determined by equation (2).

The following examples are provided to further illustrate the beneficial effects of the present invention.

FIG. 1 shows an image and its three regions A, B and C, and FIG. 2 shows the region A, B and C color histogram, where N is_color512. The histogram distance of the region A, B calculated by the formula (1) is 0.4912, the histogram distance of the region A, C is 0.0795, and the histogram distance of the region B, C is 0.9635. From this, we can judgeThe colors of the regions A, C are close together and can be merged into one region. This is consistent with the clustering results of the human eye.

On the other hand, if the conventional Bhattacharya distance is used, the histogram distance of the region A, B is 0.9995, the histogram distance of the region A, C is 0.6472, and the histogram distance of the region B, C is 0.9993. The histogram distance of the region A, C is 0.6472, indicating that the two regions are not similar in color and should not be merged into one region, which is clearly not consistent with the classification result of the human eye.

Claims

1. An image region merging method, characterized by:

wherein,

where p (k) is a color histogram of the image area a, and q (k) is a color histogram of the image area B, where k is 0,1, …, N_color-1，N_colorIndicates the number of color quantization regions used for the color histogram, [ A ]₀,A₁]A support region of p (k) [ B ]₀,B₁]Support region for q (k), D (p, q) represents the distance between the two image region color histograms: 1) if region a contains all colors that appear in region B, then the histogram distance between A, B is 0; 2) if region B contains all the colors that appear in region a, then the histogram distance between A, B is also 0; 3) if region a does not contain any color in region B, then the histogram distance between A, B is 1; 4) if not the case of 1), 2), 3), the histogram distance between A, B is