CN107247953B

CN107247953B - Feature point type selection method based on edge rate

Info

Publication number: CN107247953B
Application number: CN201710389384.4A
Authority: CN
Inventors: 林秋华; 田敏
Original assignee: Dalian University of Technology
Current assignee: Dalian University of Technology
Priority date: 2017-05-31
Filing date: 2017-05-31
Publication date: 2020-05-19
Anticipated expiration: 2037-05-31
Also published as: CN107247953A

Abstract

A feature point type selection method based on edge rate belongs to the field of computer vision. According to the invention, before feature point detection, structural information detection and classification are carried out on the image, a basis is provided for selecting a feature point type (spot or corner) suitable for the image, and the problem that matching performance is reduced and even fails due to unsuitable feature points is solved. For an image to be matched, calculating the edge of the image by using a Canny edge detection algorithm, then calculating the edge rate, and finally classifying the image according to the relationship between the edge rate and the high and low thresholds: if the edge rate is greater than the high threshold, the image structure information is very obvious, and the corner point characteristic is adopted; if the edge rate is less than the low threshold value, the image structure information is very unobvious, and the spot characteristics are adopted; if the edge rate is between the high and low thresholds, the image features are not obvious, and either the blobs or the corners can be used. The method is quick and effective, and can realize image matching with better performance by combining with the conventional image matching algorithm.

Description

Feature point type selection method based on edge rate

Technical Field

The invention relates to the field of computer vision, in particular to a feature point type selection method based on an edge rate.

Background

In the existing image matching algorithm, the method based on local feature points is most widely applied. The spots and the corners are two typical feature points, and a plurality of excellent algorithms such as SIFT and FAST provide technical support for the detection of the spots and the corners. In fact, the performance of two feature points is different for different images. The corner points are suitable for images with obvious structural information, and because the images contain more edges, intersection angles and the like, more corner points can be detected. The blobs are suitable for the image with inconspicuous structural information, and more blobs can be detected at this time, but the number of the detected corner points is less, which may cause the failure of image matching. At present, in image matching application, people often use a certain feature point detection algorithm fixedly, for example, some adopt an SIFT algorithm, namely, a speckle detection mode, and some adopt an FAST algorithm, namely, a corner detection mode. Some applications do not select a feature point category, but select a feature point category in the following two manners: one is prior information on structural characteristics of the image before image matching, and spots or corners are selected accordingly; and the other method is to carry out SIFT and FAST matching test on a small part of images, match all the images by using an algorithm with better selectivity, namely, select the feature points through a pre-matching test.

However, in practical applications such as robot visual navigation, the captured images are in a sequence, and the scene changes continuously, so the structural information of the images also changes continuously. In this case, if fixed spots or corners are used, matching may fail due to inadequate image characteristics. For example, in the Mikolajczyk image library (see, MIKOLAJCZYK, SCHMID C. application evaluation of local descriptors, IEEE Transactions on Pattern analysis and Machine analysis, vol.27, no 10, pp.1615-1630,2005), the pit image structure information is significant, and the number of detectable FAST corner points (FIG. 1(b)) is much larger than the SIFT spot points (FIG. 1 (a)). In contrast, the bark images lack structural information, and the number of detectable FAST corners (fig. 2(b)) is much smaller than the number of SIFT spots (fig. 2 (a)). Moreover, corner points detected by the bark images cannot be correctly matched, and after the false matching is eliminated by RANSAC, the number of correct matching points is equal to zero. Therefore, it is necessary to intelligently detect the structural information of each image before matching, and if the structural information is very obvious, the angular point feature is adopted; if the structural information is not obvious, adopting a spot characteristic; in other cases, both may be used, i.e. without the need for a handover algorithm.

Disclosure of Invention

The invention provides a feature point type selection method based on edge rate, which is used for detecting and classifying structural information of an image before feature point detection, providing basis for selecting a feature point type (spot or corner) suitable for the image and solving the problem of reduced matching performance and even failure caused by unsuitable feature points.

The technical scheme of the invention is that for an image to be matched, a Canny edge detection algorithm (see CANNYJ.A. computational algorithm to edge detection. IEEE transfer on pattern analysis and Machine analysis, vol.8, No.6, pp.679-698,1986) is used for calculating the edge of the image, the edge rate and the high and low threshold values thereof are defined, the edge rate is calculated according to the edge detection result, and the image is classified according to the relationship between the edge rate and the high and low threshold values: if the edge rate is greater than the high threshold, the structural information of the image is very obvious, and the angular point characteristic is adopted; if the edge rate is smaller than the low threshold value, the structural information of the image is very unobvious, and the spot characteristics are adopted; if the edge rate is between the high and low thresholds, the image features are not obvious, and both the spots and the corners can be used, the existing feature point detection algorithm can be continuously used without switching.

The method comprises the following specific steps:

firstly, Gaussian filtering is carried out on an input image to eliminate noise. The expression for gaussian filtering of an image is:

S(x,y)＝G(x,y；σ)*I(x,y) (1)

in the formula (I), the compound is shown in the specification,

the standard deviation is a Gaussian function with sigma, (x, y) is the coordinates of image pixel points, I (x, y) is an original gray image, S (x, y) is an image after filtering, and x is convolution operation. Assuming that the size of the gaussian smoothing window (also called convolution kernel) is s × s, and s is odd, the relation between σ and s is:

secondly, setting a high threshold T1 and a low threshold T2, and performing edge detection on the image by using a Canny algorithm to obtain a binary image, wherein the gray value of an edge point is 255, and the gray value of a non-edge point is 0;

thirdly, calculating the number of edge points, namely counting the number of pixel points with the gray value of 255 in the binary image obtained in the second step, and recording the number as n;

and fourthly, solving the edge rate. The edge ratio is defined as follows:

in the formula, n is the number of edge points, w is the width of an image, h is the height of the image, a is the width of an image edge pixel, a row and a column pixels are removed from the upper, lower, left and right sides of the image in the formula (3), and a is 1-3;

and fifthly, classifying the images according to the relation between the edge rate and the threshold value. The threshold values comprise a high threshold value H _ TH and a low threshold value L _ TH; when R > H _ TH, classifying the image as a very clear type of structural information; otherwise, comparing the edge rate with a low threshold, and classifying the image into a type with very inconspicuous structural information when R < L _ TH; otherwise, classifying the image as an unobtrusive feature;

and sixthly, selecting the type of the feature point. Detecting the corner characteristic of the image with very obvious structural information; for an image with very inconspicuous structural information, adopting spot characteristic detection; for an image with an inconspicuous characteristic, a spot or a corner can be used.

It should be noted that the high threshold and the low threshold of the edge rate are determined by combining the gaussian smoothing window size and by an image library experiment test. Table 1 shows the edge rate thresholds when the image library selects all images in the Mikolajczyk image library, and the gaussian smoothing window sizes are respectively selected to be 7 × 7, 9 × 9, and 11 × 11. The gaussian smoothing windows are different in size, and the high and low thresholds will be different.

TABLE 1 edge rate high and low thresholds for three Gaussian smoothing window sizes

The method has the advantages that in the image matching application based on the local feature points, a basis is provided for selecting the spot or corner features, so that not only can the matching failure caused by improper feature point selection (such as the image matching failure caused by the application of corner points in a bark image) be avoided, but also the performance of the image matching algorithm is better because the proper feature points are selected. And the speed of edge detection and the speed of edge rate calculation are both high, so that the method is quick and effective, and can realize image matching with better performance by combining with the conventional image matching algorithm.

Drawings

FIG. 1 is a comparison of spots and corners (both indicated by white dots) detected in a boat image. Wherein (a) is a spot; (b) being the corner points.

Figure 2 is a comparison of spots and corners (both indicated by white dots) detected in a bark image. Wherein (a) is a spot; (b) being the corner points.

Fig. 3 is a flow chart of the present invention.

Detailed Description

An embodiment of the present invention is described in detail below with reference to the accompanying drawings.

There is a flat image with a resolution of 850 × 680. The flow of image classification and feature point selection using the present invention is shown in fig. 3.

Firstly, performing Gaussian filtering on an input image by using a formula (1); wherein, the size of the gaussian smoothing window is selected to be 9 × 9, and the standard deviation σ is obtained to be 1.85 according to the formula (2);

secondly, setting a high threshold T1 to be 150 and a low threshold T2 to be 50, carrying out edge detection on the image by using a Canny algorithm to obtain a binary image, wherein pixel points with the gray value of 255 are edge points;

and thirdly, calculating the number of edge points. Counting the number of pixel points with the gray value of 255 in the binary image obtained in the second step to obtain n which is 43692;

fourthly, taking 3 to obtain the edge rate:

and fifthly, judging the image type according to the relation between the edge rate and the threshold value. Because the size of the gaussian smoothing window is 9 × 9, looking up table 1, the high threshold H _ TH is 0.045, and the low threshold L _ TH is 0.025. Knowing that the edge rate R of the image is 0.077, see the fifth step, so that R > H _ TH, the image is determined to be an image with very clear structure information;

and sixthly, selecting the type of the feature point. And for the image with very obvious structural information, detecting the feature points by adopting the corner point features.

Claims

1. A feature point type selection method based on edge rate is characterized by comprising the following steps:

firstly, carrying out Gaussian filtering on an input image to eliminate noise; the expression for gaussian filtering of an image is:

S(x,y)＝G(x,y；σ)*I(x,y) (1)

in the formula (I), the compound is shown in the specification,

the method is characterized in that the method is a Gaussian function with standard deviation of sigma, (x, y) is coordinates of image pixel points, I (x, y) is an original gray image, S (x, y) is an image after filtering, and x is convolution operation; assuming that the size of the gaussian smoothing window is s × s, and s is an odd number, the relationship between σ and s is:

fourthly, calculating the edge rate; the edge ratio is defined as follows:

in the formula, n is the number of edge points, w is the width of an image, h is the height of the image, and a is the width of an image edge pixel, and the row and column a of pixels are removed from the upper, lower, left and right sides of the image by the formula (3);

fifthly, classifying the images according to the relation between the edge rate and the threshold value; the threshold values comprise a high threshold value H _ TH and a low threshold value L _ TH; when R > H _ TH, the image is classified as a very clear type of structural information; otherwise, comparing the edge rate with a low threshold, and classifying the image into a type with very inconspicuous structural information when R is less than L _ TH; otherwise, classifying the image as an unobtrusive feature;

sixthly, selecting the type of the feature point; detecting the corner characteristic of the image with very obvious structural information; for an image with very inconspicuous structural information, adopting spot characteristic detection; for the image with the unobvious features, the detection of the features of the spots or the corners can be adopted.

2. The method according to claim 1, wherein the method comprises: the gaussian smoothing window size and the edge rate threshold are selected as shown in table 1:

。

3. The method according to claim 1 or 2, wherein the method comprises: the width a of the image edge pixel in the formula (3) is 1-3.