CN110705568B

CN110705568B - Optimization method for image feature point extraction

Info

Publication number: CN110705568B
Application number: CN201910882008.8A
Authority: CN
Inventors: 马书香
Original assignee: Wuhu Institute of Technology
Current assignee: Wuhu Institute of Technology
Priority date: 2019-09-18
Filing date: 2019-09-18
Publication date: 2022-07-26
Anticipated expiration: 2039-09-18
Also published as: CN110705568A

Abstract

The invention discloses an optimization method for extracting image characteristic points, which comprises the steps of inputting an image; calculating image gradient and second-order gradient; calculating an image gradient factor P (m, n) and obtaining a P (T) area; establishing a Gaussian scale space and a Gaussian difference space according to the P (T) region; carrying out extreme point detection according to P (m, n); and generating a characteristic point descriptor. Gradient image from image

And second order gradient image

Calculating factor P according to certain rule along longitudinal and transverse directions _i.j (m, n) and obtaining a P (T) area, judging whether to participate in Gaussian scale space construction and feature point extraction according to the P (T) area, omitting a large amount of calculation amount during Gaussian scale space construction and feature point extraction, and optimizing the algorithm in the feature point extraction stage and improving the algorithm efficiency due to small extra added calculation amount.

Description

Optimization method for image feature point extraction

Technical Field

The invention relates to the technical field of image processing, in particular to an optimization method for extracting image feature points.

Background

SIFT, Scale-invariant feature transform (Scale-invariant feature transform), is a description used in the field of image processing. The description has scale invariance, can detect key points in the image and is a local feature descriptor.

The image registration technology based on the image characteristics is widely applied to the fields of image processing, computer vision, pattern recognition and the like. The SIFT algorithm has strong robustness in various scenes, so that the SIFT algorithm becomes a research hotspot of an image registration technology based on image features and is widely applied.

However, the applicant has found that the conventional image feature point extraction has at least the following problems:

the SIFT algorithm takes the largest time when the feature points are extracted and are constructed in a gaussian scale space through a statistical algorithm, in the process, the algorithm constructs O × L images, for example, for an image of 1920 × 1080, the constructed images are about 20, and the larger the pixel size of the image, the more the images are. In the process, Gaussian operation is performed on all points in the image scale space, the distribution of the image feature points is not uniformly distributed in the image, and no feature point exists in a flat area or a pixel decreasing area in the image. Therefore, the SIFT algorithm has the defects of high complexity, large calculation amount and poor real-time property when the feature points are extracted.

Disclosure of Invention

In view of this, the invention aims to provide an optimization method for extracting image feature points, which solves the defects of high complexity, large calculation amount and poor real-time performance of an SIFT algorithm in feature point extraction.

Based on the above purpose, the present invention provides an optimization method for image feature point extraction, which includes:

inputting an image;

calculating image gradient and second-order gradient;

calculating an image gradient factor P (m, n) and obtaining a P (T) area;

establishing a Gaussian scale space and a Gaussian difference space according to the P (T) region;

carrying out extreme point detection according to the P (T) area;

and generating a characteristic point descriptor.

The input image comprises a Gaussian operation of a first image in a scale space on an initial image, a Gaussian scale space structure of an optimization method is carried out on the basis of the image, and the image after the Gaussian operation is preprocessed.

The performing image gradient and second order gradient calculations comprises:

calculating the gradient values of the image in the longitudinal and transverse directions, wherein the calculation result of each pixel point is

Wherein j is the row of the image, and i is the pixel coordinate position of the j row of the image;

calculating a second-order gradient value of the image in the longitudinal direction and the transverse direction, wherein the result of calculation of each pixel point is

The calculating the image gradient factor P (m, n) and obtaining the P (T) region comprises:

initial value P of gradient factor _i,j (m,n)＝P _0,0 (0,0), where K is i (for the starting position of j row, K is 0, and for the pixel points at other positions, K is i);

judging whether P is present _i,j (dx,) and P _i-1,j (dx,) symbols are coincident and | P _i,j (d' x,) | or | P _i-1,j (d' x,) | is less than S;

if so, then P _i,j The gradient factor value m +1, and judging the calculation condition of j rows of pixels of the image;

otherwise, it is determined that K is 0.

The determination that K is 0 includes:

if K is equal to 0, then P _i,j Gradient factor value is denoted m, inverse direction P _i,j To P _0,j Sequentially subtracting 1 from the m value of the image, and judging the calculation condition of j rows of pixels of the image;

if K is not equal to 0, then P _i-m/2,j Gradient factor values are denoted m, P _i-m/2,j To P _i,j Sequentially subtracting 1 from the m value of (b), and judging the calculation condition of the pixels of j rows of the image.

The judgment of the j line calculation condition of the image comprises the following steps:

judging whether the calculation of the j line of the image is finished, if so, continuing to calculate the j ═ j +1 of the next line;

otherwise, calculate P _i+1,j I +1, and determines whether P is present _i,j (dx,) and P _i-1,j (dx,) symbols are coincident and | P _i,j (d' x,) | or | P _i-1,j (d' x,) | is less than S.

Judging whether the calculation of the j line of the image is finished, if so, continuing to calculate the j ═ j +1 of the next line comprises the following steps:

calculating P _i,j A gradient factor value n;

outputting an image gradient factor P _i.j (m,n)。

The calculating of the image gradient factor P (m, n) and obtaining the P (t) region comprises:

according to P of the obtained image of the order _i.j (m, n) value, each pixel point is processed in the construction of scale spaceJudging m and n values before performing Gaussian operation, and performing Gaussian operation to participate in the construction of a scale space after the m and n values are smaller than the T value;

wherein, the T value is larger than half of the field size required by the calculation of the characteristic point descriptor; only one P per image level is required _i.j The value of (m, n) is sufficient.

The extreme point detection according to the P (T) area comprises the following steps: when extreme point extraction is carried out on the Gaussian difference space images, the extreme value extraction area of each image is as follows: the P (T) region of the level image. And sequentially obtaining extreme points.

From the above, the optimization method for extracting the image feature points provided by the invention comprises the steps of inputting an image; calculating image gradient and second-order gradient; calculating an image gradient factor P (m, n) and obtaining a P (T) area; establishing a Gaussian scale space and a Gaussian difference space according to the P (T) region; carrying out extreme point detection according to the P (T) area; and generating a characteristic point descriptor. Gradient image from image

And second order gradient image

Calculating factor P according to certain rule along longitudinal and transverse directions _i.j (m, n), and further obtaining a P (T) region according to a certain rule to judge whether to participate in Gaussian scale space construction and feature point extraction, so that a large amount of calculation amount in Gaussian scale space construction and feature point extraction is omitted, the extra calculation amount added by the method is small, the algorithm is optimized in the feature point extraction stage, and the algorithm efficiency is improved.

Drawings

FIG. 1 is a schematic block diagram of a flow chart of an optimization method for extracting image feature points according to an embodiment of the present invention;

fig. 2 is a block diagram illustrating an exemplary flow of image gradient factor calculation in an optimization method for extracting image feature points according to an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail with reference to the following embodiments and the accompanying drawings.

As one embodiment of the present invention, as shown in fig. 1, a method for optimizing image feature point extraction includes:

step 101, inputting an image;

the step may further include: and Gaussian filtering preprocessing, specifically, performing Gaussian operation on the first image in the scale space on the initial image, and performing operation of an optimization algorithm on the basis of the image.

102, calculating image gradient and second-order gradient;

the step may further include:

step 201, calculating gradient values of the image in two directions of vertical and horizontal directions, wherein the calculation result of each pixel point is

The pixel points of the i, j coordinates in the vertical and horizontal directions on the image are P (i, j), P _i,j (dx, dy) is the differential of P (i, j) in the longitudinal and lateral directions, dx is the differential of P (i, j) in the x direction

dy is the differential of P (i, j) in the y direction

P _i,j (d ' x, d ' y) is the second order differential of P (i, j) in the longitudinal and lateral directions, and d ' x is the second order differential of P (i, j) in the x direction

d' y is the second differential of P (i, j) in the y direction

Step 103, calculating an image gradient factor P (m, n) and obtaining a P (T) area;

the steps may also include:

step 301, based on the operation in step 201, calculating a gradient factor P according to a predetermined rule along the vertical and horizontal directions from the operation result in step 201 _i.j (m, n), starting from a certain vertex of the image, processing the image pixel by pixel to obtain P _i.j (m, n), the initial values of m and n are marked as 0, and the calculation methods of m and n are the same.

Step 301 includes:

step 3011, start line by line scanning from the first line of the image, P _i,j (m,) value calculation method: if P _i,j (dx,) and P _i-1,j (dx) there is no change in sign (sign is also considered unchanged when a gradient value of 0 is encountered), and P _i,j (d' x,) or P _i-1,j If (d' x,) are all less than S value, then P _i,j (m+1,)。

If the sign of the gradient value changes or the second-order gradient value is larger than S, P is _i,j The m value of (m,) is marked as 0, then the m value of the previous m/2 pixel point changes, P _i-m/2,j The value of (m,) is marked m, P _i-m/2,j To P _i,j The m value of (a) is sequentially decreased by 1.

The step may further include: if it is the value of m calculated from the beginning of each line, P _i,j Gradient factor value is denoted m, inverse direction P _i,j To P _0,j Sequentially subtracting 1 from the m value of the image, and judging the calculation condition of j rows of pixels of the image; a new m is calculated.

Step 3012, go back to P _i+1.j And (m,) marking the value m as 0, continuing the calculation in the step 3011 until the pixels in the row are finished.

Step 3013, scanning the next row of pixel points of the image, and repeating the operation methods of step 3011 and step 3012;

step 3014, according to P in step 3011 _i,j (m,) value calculation method, calculating P _i,j (, n) obtaining P for the order image _i.j (m, n) values.

Wherein, according to P of the obtained image of the order _i.j When the values of m and n are both smaller than the value of T, the part of the region is marked as P (T), and the P (T) region is subjected to Gaussian operation to participate in the construction of a scale space;

wherein, the value of T is more than half of the domain size required by the feature descriptor calculation; only one P per image level is required _i.j The value of (m, n) is sufficient.

Step 104, establishing a Gaussian scale space and a Gaussian difference space according to the P (T) region;

step 105, performing extreme point detection according to the P (T) area;

the step may further include: when extreme point extraction is carried out on the Gaussian difference space image, the extreme extraction area of each image is a P (T) area. And sequentially obtaining extreme points.

And 106, generating a characteristic point descriptor.

Embedding the optimization method in a characteristic point extraction stage of an SIFT original algorithm, preprocessing an initial image before constructing a scale space, identifying a pixel part which does not participate in the construction of the scale space, not performing Gaussian operation processing during subsequent construction of the scale space, and skipping the areas during a characteristic point detection stage of the algorithm so as to improve the algorithm efficiency;

FIG. 2 is an exemplary flow diagram of image gradient factor calculation;

the method comprises the following steps: step 10, inputting an image;

step 11, Gaussian filtering pretreatment;

step 12, image gradient value P _i,j (dx, dy) and a second order gradient value P _i,j (d 'x, d' y) calculation;

step 13, initial value P of gradient factor _i,j (m,n)＝P _0,0 (0,0), where K is i, (for the pixel at the start position of j rows, K is 0, and for the pixel at the other positions, K is i);

step 14, determine whether P is present _i,j (dx,) and P _i-1,j (dx,) symbols are coincident and | P _i,j (d' x,) | or | P _i-1,j (d' x,) | is less than S;

step 15, if yes, P _i,j Gradient factor value m +1, and go to step 19;

step 16, otherwise, judging whether K is equal to 0;

step 17, if K is equal to 0, P _i,j Gradient factor value is denoted m, inverse direction P _i,j To P _0,j Decreasing the m value by 1 in sequence, and proceeding to step 19;

step 18, if K is not 0, P _i-m/2,j Gradient factor values are denoted m, P _i-m/2,j To P _i,j Decreasing the m value by 1 in sequence, and proceeding to step 19;

step 19, judging whether the calculation of the j line of the image is finished; if yes, go to step 21, otherwise go to step 20;

step 20, calculating P _i+1,j I +1 and proceeds to step 14;

step 21, continuing to calculate j ═ j +1 in the next row;

step 22, calculate P _i,j A gradient factor value n;

step 23, outputting the image gradient factor P _i.j (m,n)。

The method has small added algorithm operation amount, and improves the SIFT feature point extraction speed.

Gradient image from image

And second order gradient images

Calculating factor P according to certain rule along longitudinal and transverse directions _i.j (m, n) and obtaining a P (T) area, judging whether the pixel point area participates in the Gaussian scale space structure and the feature point extraction according to the P (T) area, omitting a large amount of computation load during the Gaussian scale space structure and the feature point extraction, and the method has small extra added computation load, optimizes the algorithm in the feature point extraction stage and improves the algorithm efficiency.

Those of ordinary skill in the art will understand that: the discussion of any embodiment above is meant to be exemplary only, and is not intended to intimate that the scope of the disclosure, including the claims), is limited to these examples; within the idea of the invention, also features in the above embodiments or in different embodiments may be combined, steps may be implemented in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in detail for the sake of brevity.

While the present invention has been described in conjunction with specific embodiments thereof, many alternatives, modifications, and variations of these embodiments will be apparent to those skilled in the art in light of the foregoing description. E.g., dynamic RAMDRAM)) may use the discussed embodiments.

The embodiments of the invention are intended to embrace all such alternatives, modifications and variances that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalents, improvements, and the like that may be made without departing from the spirit or scope of the present invention are intended to be included within the scope of the present invention.

Claims

1. An optimization method for extracting image feature points is characterized by comprising the following steps:

inputting an image;

calculating image gradient and second-order gradient;

calculating an image gradient factor P (m, n) and obtaining a P (T) area;

carrying out extreme point detection according to the P (T) area;

generating a characteristic point descriptor;

initial value P of gradient factor _i,j (m,n)＝P _0,0 (0,0)，K＝i；

Wherein j is a row of the image, and i is a pixel coordinate position of the j row of the image;

if so, then P _i,j The gradient factor value m +1, and the calculation condition of j rows of the image is judged;

otherwise, the case where K is 0 is determined.

2. The method as claimed in claim 1, wherein the input image comprises a gaussian operation of a first image in a scale space on an initial image, and the optimization algorithm is performed based on the first image to preprocess the gaussian filtering.

3. The method of claim 1, wherein the performing image gradient and second order gradient calculations comprises:

4. The method according to claim 1, wherein the determining that K-0 holds true comprises:

if K is not equal to 0, then P _i-m/2,j Gradient factor values are denoted m, P _i-m/2,j To P _i,j Sequentially subtracting 1 from the m value of the image, and judging the calculation condition of j lines of the image.

5. The method according to claim 1 or 4, wherein the determining j rows of pixel calculation conditions of the image comprises:

judging whether the calculation of the pixels in the j line of the image is finished, if so, continuing to calculate the j line as j + 1;

6. The method as claimed in claim 5, wherein the determining whether the calculation of j rows of the image is finished, and if yes, continuing to calculate j ═ j +1 in the next row comprises:

calculating P _i,j A gradient factor value n;

output image gradient factor P _i.j (m，n)。

7. The method of claim 1, wherein the calculating an image gradient factor P (m, n) and obtaining a region P (t) comprises:

according to P of the obtained image _i.j The (m, n) values are judged before Gaussian operation is carried out on each pixel point during construction of the scale space, and the Gaussian operation is carried out to participate in construction of the scale space after the m and n values are smaller than the T value;

8. The method of claim 1, wherein the performing extreme point detection according to the p (t) region comprises: when extreme point extraction is carried out on the Gaussian difference space image, the extreme extraction area of each image is as follows: and obtaining extreme points in the P (T) area of the image in sequence.