CN105913068B

CN105913068B - It is a kind of for describing the multi-dimensional direction gradient representation method of characteristics of image

Info

Publication number: CN105913068B
Application number: CN201610268754.4A
Authority: CN
Inventors: 石柱国
Original assignee: Beijing Yisa Technology Co Ltd
Current assignee: ISSA Technology Co Ltd
Priority date: 2016-04-27
Filing date: 2016-04-27
Publication date: 2018-12-25
Anticipated expiration: 2036-04-27
Also published as: CN105913068A

Abstract

The present invention provides a kind of for describing the multi-dimensional direction gradient representation method of characteristics of image, comprising: defines the matrix of A × A size respectively centered on pixel each in image, A therein takes 3~30 odd number；The gradient of x and y both direction is calculated for central pixel point in each matrix of definition；The gradient calculated result obtained based on each matrix is weighted, x direction gradient and y direction gradient that each matrix is directed to its central pixel point are obtained；Again in the block of the N × N further defined, the range of 0-180 degree is divided into 5-9 unit of equal angular, the gradient weight in each unit is counted using linear interpolation；Centered on each pixel in image, the matrix of a M × M is done, the maximum value vmax and minimum value vmin of gradient are found in this matrix；Recalculate gradient weight.The block zone algorithm of method utilization orientation gradient of the invention, while the linear interpolation in angle is integrated, object features can be described well, while ensure that calculating speed.

Description

It is a kind of for describing the multi-dimensional direction gradient representation method of characteristics of image

Technical field

The invention belongs to Computer Image Processing fields, for accurately describing image object feature.

Background technique

Local image characteristics description be computer vision a basic research problem, find image in corresponding points with And there is important role in object features description.It is the basis of many methods, thus be also current vision research in one A hot spot.It also has a wide range of applications simultaneously, for example, is carrying out three-dimensional reconstruction using two-dimensional images, is restoring field In the application of scape three-dimensional structure, basic point of departure is that have a reliable image to correspond to point set, and automatically set up figure Reliable corresponding relationship as between points usually all relies on outstanding local image characteristics description.Compare again Such as, very popular at present and practicable method first is that based on local feature, due to feature in object identification Locality so that object identification can handle block, the more complicated situation such as complex background.

The key problem of local image characteristics description is invariance (robustness) and ga s safety degree.However, Feature Descriptor The power of ga s safety degree be often contradictory with its invariance, that is to say, that a feature with numerous invariance describes Son, the ability that it distinguishes local image content are just slightly weak；And if a spy for being very easy to distinguish different local image contents Sign description, its robustness are often relatively low.And on the other hand, if we by count local image grey level histogram come Situations such as progress feature description, this describing mode has stronger invariance, rotates variation for local image content Compare robust, but separating capacity is weaker, such as two topographies that grey level histogram is identical but content is different cannot be distinguished Block.So an outstanding Feature Descriptor should not only have very strong invariance, should also have very strong ga s safety degree.

In existing many local image characteristics description, SIFT (Scale Invariant Feature Transform) be it is wherein most widely used, it was put forward for the first time in 1999, until 2004 improve.Since SIFT is to ruler Degree, rotation and certain visual angle and illumination variation etc. image changes all have invariance, and SIFT has very strong distinguish Property, since it is proposed, applied in object identification, wide baseline image matching, three-dimensional reconstruction, image retrieval quickly, Local image characteristics description has also obtained more extensive concern in computer vision field, and it is special to have emerged in large numbers large quantities of each tools The local image characteristics of color describe son.SURF (Speeded Up Robust Features) is the modified version to SIFT, it Using Haar small echo come the gradient operation in approximate SIFT method, while quickly being calculated using integral diagram technology, SURF's Speed is 3-7 times of SIFT, it is suitable with the performance of SIFT in most cases, therefore it is answered in many applications With especially to the demanding occasion of runing time.ASIFT (Affine SIFT) is obtained under all imaged viewing angles by simulating Image carry out characteristic matching, the case where visual angle change can be handled well, the images match under especially big visual angle change. MROGH (Multi-support Region Order-based Gradient Histogram) is then on feature convergence strategy To seek to innovate, local image characteristics before describe son, and feature convergence strategy is all based on the geometric position put in neighborhood, And MROGH carries out feature convergence based on the gray scale sequence of point.LBP (Local Binary Pattern, local binary patterns) is one Kind it is used to describe the operators of image local textural characteristics, first by T.Ojala,With D.Harwood 1994 It is proposed occur many mutation, such as CLBP later year, LBPHF etc. is used for image local texture feature extraction.LBP has succeeded Applied to Face datection, lip reading identification, expression is detected, and dynamic texture etc. field is constant with rotational invariance and gray scale The significant advantages such as property.

The development trend of local image characteristics recent years description is: quick, low storage.The two trend make part Image feature descriptor can quickly play a role in real time, in large-scale application, and be conducive to do many applications in one's hands Machine is developed up, and true is applied to computer vision technique in the world around us.The above-mentioned prior art The middle method that characteristics of image is described using operators such as SIFT and LBP is in calculating speed, resource consumption and to feature description In compactness, all there are still insufficient and defects.In order to which further satisfaction is quick and low the two demands of storage, it is necessary to research and develop A kind of new new image representation method.

Summary of the invention

It is an object of the invention to: provide a kind of calculating speed faster, that resource consumption is lower, feature describes compactness is higher New image representation method.

Above-mentioned purpose of the invention is achieved through the following technical solutions:

It provides a kind of for describing the multi-dimensional direction gradient representation method of characteristics of image, comprising:

1) matrix of A × A size is defined respectively centered on pixel each in image, A therein takes 3~30 surprise Number；

2) gradient of x and y both direction is calculated for central pixel point in each matrix that step 1) defines；

3) the gradient calculated result obtained to step 2) based on each matrix is weighted, and is obtained each matrix and is directed to The x direction gradient and y direction gradient of its central pixel point, i.e., each pixel correspondence obtains the gradient in 1 direction x in image The gradient value of value and 1 direction y；

4) block of N × N size is defined, the N takes positive integer；In the block of definition, by the range of 0-180 degree point For 5-9 unit of equal angular, the gradient weight in each unit is counted using linear interpolation；

5) centered on each pixel in image, the matrix of a M × M is done, wherein M is that N described in step 4) takes The integral multiple of value finds the maximum value vmax and minimum value vmin of gradient in this matrix；Gradient weight is recalculated, is calculated Method are as follows: (v-vmin)/(vmax-vmin).

In method of the invention, A × A matrix described in step 1), A is preferably 5~15；More preferable 5.

In method of the invention, x and y both direction is calculated for central pixel point in each matrix described in step 2) Gradient, preferred calculation method is: being that symmetry axis calculates the pixel value between every group of symmetric points with column where central pixel point Difference；The margin of image element between every group of symmetric points is calculated with the behavior symmetry axis where central pixel point.

In method of the invention, weighted calculation described in step 3), preferably by defining one and the step 1) matrix The Gaussian kernel of same size completes the weighted calculation.

In method of the invention, the block of the preferred N × N size of step 4), N takes 8.

In method of the invention, the range of the 0-180 degree is preferably divided into 9 units, each unit 20 by step 4) Degree.

In method of the invention, the matrix of the preferred M × M of step 5), M takes 32.

The block zone algorithm of method utilization orientation gradient of the invention, it is possible to reduce gradient caused by local light shines unevenly Differentiation is excessive, improves the adaptability of gradient description, while integrating the linear interpolation in angle, can describe object spy well Sign, while ensure that calculating speed.

Specific embodiment

Below by way of the mode for enumerating embodiment, the following further describes the technical solution of the present invention.

Embodiment 1

It is a kind of for describing the multi-dimensional direction gradient representation method of characteristics of image, comprising the following steps:

1. the calculating of direction gradient

A) for each of image point, determine one 5 × 5 matrix, each position is in matrix with English words female mark Remember as follows:

B) central point in each matrix that step a) is defined for position such as m point calculates gradient as follows:

First calculate x direction gradient:

Gradx1=(a-e) * CX1+ (b-d) the * row of CX2//the 1st

Gradx2=(f-j) * CX3+ (g-i) the * row of CX4//the 2nd

Gradx3=(k-o) * CX5+ (l-n) the * row of CX6//the 3rd

Gradx4=(p-t) * CX7+ (q-s) the * row of CX8//the 4th

Gradx5=(u-y) * CX9+ (v-x) the * row of CX10//the 5th

Wherein CX1~CX9 is gaussian coefficient

After obtaining the gradient in this 5 directions x gradx1~grad5, then addition is weighted to this 5 values, obtains m point X direction gradient.

Similarly, the y direction gradient of m point can be obtained.

2. defining one 8 × 8 block in the picture to count direction gradient in this block, according to 0-180 degree, often 20 degree are a unit, count the weight in this 9 units.Linear interpolation is used in statistics, the gradient of a such as point is 23 degree, then (23-20)/20 × gradient value, which obtains value, is classified as 20 degree to 40 degree this sections；

3. doing one 32 × 32 matrix centered on each of image pixel, step 1 is found in this matrix The maximum value vmax and minimum value vmin for the gradient being calculated.Recalculating gradient weight is (v-vmin)/(vmax- Vmin), excessive to reduce gradient difference alienation caused by local light shines unevenly, improve the adaptability of gradient description.

Claims

1. a kind of for describing the multi-dimensional direction gradient representation method of characteristics of image, comprising:

1) matrix of A × A size is defined respectively centered on pixel each in image, A therein takes 3~30 odd number；

3) the gradient calculated result obtained to step 2) based on each matrix is weighted, and obtains each matrix for wherein The x direction gradient and y direction gradient of imago vegetarian refreshments, i.e., in image the corresponding gradient value for obtaining 1 direction x of each pixel and The gradient value in 1 direction y；

4) block of N × N size is defined, the N takes positive integer；In the block of definition, the range of 0-180 degree is divided into phase With 5-9 unit of angle, the gradient weight in each unit is counted using linear interpolation；

5) centered on each pixel in image, the matrix of a M × M is done, wherein M is N value described in step 4) Integral multiple finds the maximum value vmax and minimum value vmin of gradient in this matrix；Recalculate gradient weight, calculation method Are as follows: (v-vmin)/(vmax-vmin).

2. method described in claim 1, it is characterised in that: A × A matrix described in step 1), A are 5~15.

3. method described in claim 1, it is characterised in that: A × A matrix described in step 1), A 5.

4. method described in claim 1, it is characterised in that: be directed in each matrix that step 1) defines described in step 2) Central pixel point calculates the gradient of x and y both direction, and calculation method is: being that symmetry axis calculating is every with column where central pixel point Margin of image element between group symmetric points；The pixel value between every group of symmetric points is calculated with the behavior symmetry axis where central pixel point Difference.

5. method described in claim 1, it is characterised in that: weighted calculation described in step 3) is by defining one and step The Gaussian kernel of rapid 1) the described matrix same size completes the weighted calculation.

6. method described in claim 1, it is characterised in that: the block of N × N size, N described in step 4) take 8.

7. method described in claim 1, it is characterised in that: step 4) is that the range of the 0-180 degree is divided into 9 lists Position, 20 degree of each unit.

8. method described in claim 1, it is characterised in that: the matrix of M × M of step 5), M take 32.