CN106022342A - Image feature extraction method based on KAZE algorithm - Google Patents

Abstract

The invention discloses an image feature extraction method based on a KAZE algorithm. The problem of low implementation efficiency of the existing image feature extraction technology is solved. The method comprises the steps that a nonlinear partial differential equation is constructed; an AOS algorithm is used to solve an equation to acquire all images in a non-linear scale space; feature points are detected and sub-pixels are accurately positioned; the main direction is determined according to the local image structure of the feature points; a selected window is used to calculate the description vector of a sub-region; dimension reduction is carried out on the acquired description vector through a principal component analysis method; and finally, feature matching is carried out. According to the method, the principal component analysis method is used to reduce the dimension of a descriptor to speed up the calculation speed, which has a good undertaking performance in image feature extraction and matching; and the instantaneity and the matching rate of the algorithm are improved.

Description

A kind of image characteristic extracting method based on KAZE algorithm

Technical field

The invention belongs to computer vision and image processing field, be specifically related to a kind of image based on KAZE algorithm special Levy extracting method.

Background technology

In recent years, along with the progress of image processing techniques, image characteristics extraction has obtained significant progress.Characteristics of image Extractive technique can be applied at aspects such as daily life, industrial or agricultural detection, biotechnology, medical science detections.Feature extraction It is a primary computing in image procossing, first calculation process namely an image carried out.Characteristics of image Global characteristics and local feature can be divided into.Global characteristics refers mainly to the variance of image, color histogram etc..And local Feature is more focused on the feature that image local occurs, local feature can be understood as stable existence and has good discrimination The point of character.Play a significant role at aspects such as prospect background differentiation, object identifications.

Local invariant feature refers to the detection of local feature or describes the various changes to image, such as, becomes for visual angle The invariance changed, to the invariance of dimensional variation, to rotationally-varying invariance, to invariance of shape etc..Base Application in local feature comprises three basic steps: detects, describe and mates.Good local image characteristics should have Having speed fast, feature description has preferable robustness to yardstick, illumination, rotation, and feature description dimension is low simultaneously, easily In realizing the feature that Rapid matching is the most real-time.The feature extracted from image can form multiple vector, two The similarity between them can be calculated by the measurement degree of certain distance metric or similarity between image.

How to extract from original image and there is the study hotspot that the characteristics of image of relatively strong representation ability is image procossing. Local shape factor in early days has Harris corner detection operator, and " angle point " is often detected in demarcation line, object Boundary, texture is than stronger part.Traditional SIFT, SURF operator for yardstick, illumination, visual angle, The change rotated has stronger robustness, but relatively low for image border and details reservation degree, fuzzy serious, and Nonlinear diffusion filtering employed in KAZE algorithm preferably solves this problem.KAZE algorithm is non-linear Carry out feature point detection and location on the basis of metric space, use M-Surf method to carry out characteristic point description.Anhui work Li Dan that sparetime university is learned et al. has delivered " a kind of improvement on Sichuan University's journal (the 3rd phase of volume 52 in May, 2015) KAZE feature detection algorithm " paper, this paper by improve characteristic point search strategy, utilize circle improve feature to The steps such as amount description method attempt improving the real-time of algorithm, but there is the problem that execution efficiency is on the low side.

Summary of the invention

The technical problem to be solved in the present invention is to propose a kind of image characteristics extraction scheme based on KAZE, it is possible to protecting Improve matching rate on the basis of staying real-time further, improve the execution efficiency of algorithm.

For solving above-mentioned technical problem, the technical scheme that the present invention proposes is a kind of image characteristics extraction based on KAZE Method, comprises the steps of

Step one: input picture L, constructs nonlinear partial differential equation, carries out Nonlinear diffusion filtering, recycling AOS Algorithm for Solving equation obtains all images of Nonlinear Scale Space Theory；

Step 2: feature point detection, finds spy by the Hessian local maximum after finding different scale normalization Levy a little, after finding the position of characteristic point, according to Taylor expansion, carry out sub-pix and be accurately positioned

Step 3: determine its principal direction according to the local image structure of characteristic point；

Step 4: be σ for scale parameter_iCharacteristic point, gradient image takes centered by characteristic point the window of Mouthful, and window is divided into subregion, calculate the description vectors of subregion；

Step 5: the description vectors principal component analytical method obtained is carried out dimension-reduction treatment, obtains the description after dimensionality reduction Son, then carry out characteristic matching.

Above-mentioned Nonlinear diffusion filtering is brightness of image change on different scale to be regarded by nonlinear partial differential equation Divergence for some form of flow functionBy arranging propagation function c, control to expand The degree of dissipating and classification allow it be adaptive to the partial structurtes of image, utilize AOS algorithm to obtain the institute of Nonlinear Scale Space Theory There is image.

Further, the principal component analytical method described in step 5 comprises the steps of

Step 1: input describes son；

Step 2: to describing substandard；

Step 3: structure covariance matrix, obtains characteristic vector and eigenvalue；

Step 4: arrange the matrix made new advances and screen；

Step 5: be mapped in new matrix, obtains of the description after dimensionality reduction.

Further, in step 2, at the Hessian matrix of structureMiddle searching pole During value point, in order to accelerate search speed, each pixel and its current scale, a upper yardstick, next yardstick encloses Pixel under the window of its fixing 3*3 size compares, to guarantee in metric space and two dimensional image space Extreme point detected.

In step 3, determine that characteristic point principal direction is by arranging characteristic point search radius, in circle the one of all adjoint points Rank differential value is weighted by Gauss so that big near the respective contribution of characteristic point, the respective contribution away from characteristic point is little, Differential value is regarded as vector value, and is main by move window traversal border circular areas obtaining the angle-determining of long vector by sector Direction.

In step 4, for described characteristic point, gradient image takes centered by characteristic point 24 σ_i×24σ_i's Window, and window is divided into 4 × 4 sub regions, every sub regions size is 9 σ_i×9σ_i, adjacent subregion has Width is 2 σ_iOverlap, every sub regions all uses a gaussian kernel (σ₁=2.5 σ_i) be weighted, then calculate The subregion description vectors of a length of 4: d_v=(∑ L_x, ∑ L_y, Σ | L_x|, Σ | L_y|), then be 4 by another size The Gauss window of × 4 vectorial d to every sub regions_vIt is weighted, is finally normalized.Thus obtain The description vectors of 4 × 4 × 4=64 dimension, in like manner, it is also possible to by changing window size, obtain the description of 128 dimensions to Amount.

Beneficial effect: the design that the characteristics of image figure based on KAZE that the present invention proposes extracts, compared to existing Image characteristics extraction algorithm, the program has the advantage that

(1) current feature detection algorithm in addition to KAZE is all based on linear-scale space, and KAZE is adopted Nonlinear Scale Space Theory loss in detail few, edge retains more preferably, and information retains more complete.

(2) this method uses principal component analytical method to carry out dimensionality reduction to describing son, accelerates to calculate speed, special for image Levy extraction, with mating, there is good undertaking effect, in terms of the real-time improving algorithm and matching rate, have preferable effect.

Accompanying drawing explanation

Fig. 1 is whole image characteristics extraction based on KAZE and the schematic flow sheet mated；

Fig. 2 is to describing the sub flow chart carrying out principal component analysis.

Detailed description of the invention

It is embodied as being further described in detail to the present invention in conjunction with accompanying drawing.

The present invention is a kind of to use principal component analytical method to carry out the feature extraction scheme of dimensionality reduction to describing son.Based on KAZE Image characteristics extraction algorithm, it is contemplated that principal component analytical method can retain key data composition in terms of data process On the premise of carry out effective dimensionality reduction, apply the method in describe son process on, extract feature at KAZE algorithm After point describes son, sub-dimension will be described with principal component analytical method and reduce, remove noise, improve images match efficiency. Comparing more existing design cycle and KAZE feature extraction algorithm, the present invention proposes a kind of use principal component analysis and adds The scheme of speed KAZE image characteristics extraction speed.The generation of Feature Descriptor be the final step of feature extraction be also A crucial step, each describes son and contains the information such as the position of characteristic point, gradient direction, angle, by main one-tenth The method analyzed, reduces the dimension describing son, can accelerate scheme and realize speed.The realization of the program and applying Journey is as follows:

As it is shown in figure 1, image characteristics extraction scheme based on KAZE, it is included in detail below in step:

Step one: for input picture, uses additive operator splitting algorithm (Additive Operator Splitting, AOS) Carry out Nonlinear diffusion filtering.Nonlinear diffusion filtering is to say that brightness of image is in difference by nonlinear partial differential equation Change on yardstick is considered as the divergence of some form of flow functionRecycling AOS Algorithm for Solving equation, i.e. can get all images of Nonlinear Scale Space Theory.

Step 2: feature point detection, finds spy by the Hessian local maximum after finding different scale normalization Levy a little.Hessian matrix at structureDuring middle searching extreme point, in order to accelerate to search Suo Sudu, each pixel and its current scale, a upper yardstick, around its fixing 3*3 size on next yardstick Window under pixel compare, to guarantee extreme point to be detected at metric space and two dimensional image space.Find Behind the position of characteristic point, according to Taylor expansion, carry out sub-pix and be accurately positioned

Step 3: in order to realize the rotational invariance of image, needs to determine it according to the local image structure of characteristic point Principal direction.Here the method that author uses is similar to SURF, by arranging characteristic point search radius, to all in circle The first differential value of adjoint point is weighted by Gauss, and the respective contribution of the close characteristic point being is big, corresponding away from characteristic point Contribute little, differential value is regarded as vector value, and obtains the angle of long vector with fan-shaped sliding window traversal border circular areas It is exactly principal direction.

Step 4: be σ for scale parameter_iCharacteristic point, gradient image takes centered by characteristic point 24 σ_i ×24σ_iWindow, and window is divided into 4 × 4 sub regions, every sub regions size is 9 σ_i×9σ_i, adjacent Subregion has width to be 2 σ_iOverlap.Every sub regions all uses a gaussian kernel (σ₁=2.5 σ_i) be weighted, so After calculate a length of 4 subregion description vectors: d_v=(∑ L_x, ∑ L_y, ∑ | L_x|, ∑ | L_y|), then by another Size is the Gauss window of 4 × 4 vectorial d to every sub regions_vIt is weighted, is finally normalized.So The description vectors just having obtained 4 × 4 × 4=64 dimension (in like manner can also obtain retouching of 128 dimensions by changing window size State vector).

Step 5: the description son principal component analytical method obtained is carried out dimension-reduction treatment, as in figure 2 it is shown, to description Construct covariance matrix after substandard, obtain characteristic vector and eigenvalue, filter out the characteristic vector of reservation and reflect It is mapped in new matrix, obtains of the description after dimensionality reduction, then carry out characteristic matching.Concrete operations are as follows:

(1) in actual application, should first eliminate the impact of guiding principle amount before calculating main constituent, one of conventional method is just It is that initial data is standardized, it is possible to first original description is standardized, side

Just subsequent calculations:

$x_{ij}^{*}=\frac{x_{ij}-\stackrel{\‾}{x_j}}{s_j},i=1,2,\mathrm{...},n;j=1,2,\mathrm{...},p$

Wherein:

(2) for describing submatrix, his covariance matrix is calculated:

$s_{ij}=\frac{1}{n-1}\underset{k=1}{\overset{n}{\mathrm{\Σ}}}(x_{ki}-\stackrel{\‾}{x_i})(x_{kj}-\stackrel{\‾}{x_j}),i,j=1,2,\mathrm{...},p$

(3) eigenvalue and the characteristic vector of covariance matrix are obtained；

(4) characteristic vector is arranged in matrix from top to bottom by character pair value size, takes the matrix that front k row composition is new；

(5) by maps feature vectors obtained in the previous step to new space matrix, description that dimensionality reduction is later is obtained.

Claims (6)

1. an image characteristic extracting method based on KAZE, it is characterised in that comprise the steps of

Step one: input picture L, constructs nonlinear partial differential equation, carries out Nonlinear diffusion filtering, recycles AOS algorithm Solving equation obtains all images of Nonlinear Scale Space Theory；

Step 2: feature point detection, finds characteristic point by the Hessian local maximum after finding different scale normalization, After finding the position of characteristic point, according to Taylor expansion, carry out sub-pix and be accurately positioned

Step 3: determine its principal direction according to the local image structure of characteristic point；

Step 4: be σ for scale parameter_iCharacteristic point, gradient image takes centered by characteristic point the window of, and Window is divided into subregion, calculates the description vectors of subregion；

Step 5: the description vectors principal component analytical method obtained is carried out dimension-reduction treatment, obtains of the description after dimensionality reduction, then Carry out characteristic matching.

A kind of image characteristic extracting method based on KAZE, it is characterised in that described non-linear expansion Scattered filtering is, by nonlinear partial differential equation, brightness of image change on different scale is considered as some form of flow function DivergenceBy arranging propagation function c, control diffusion and classification allows it be adaptive to image Partial structurtes, utilize AOS algorithm to obtain all images of Nonlinear Scale Space Theory.

A kind of image characteristic extracting method based on KAZE, it is characterised in that described in step 5 Principal component analytical method comprise the steps of

Step 1: input describes son；

Step 2: to describing substandard；

Step 3: structure covariance matrix, obtains characteristic vector and eigenvalue；

Step 4: arrange the matrix made new advances and screen；

Step 5: be mapped in new matrix, obtains of the description after dimensionality reduction.

A kind of image characteristic extracting method based on KAZE, it is characterised in that at structure in step 2 The Hessian matrix madeDuring middle searching extreme point, in order to accelerate search speed, each Pixel and its current scale, a upper yardstick, next yardstick is carried out around the pixel under the window of its fixing 3*3 size Relatively, to guarantee extreme point to be detected at metric space and two dimensional image space.

A kind of image characteristic extracting method based on KAZE, it is characterised in that determine in step 3 Characteristic point principal direction is by arranging characteristic point search radius, is weighted the first differential value of all adjoint points in circle by Gauss, makes Must be big near the respective contribution of characteristic point, the respective contribution away from characteristic point is little, and differential value regards as vector value, and will be with fan-shaped The angle-determining that dynamic window traversal border circular areas obtains long vector is principal direction.

A kind of image characteristic extracting method based on KAZE, it is characterised in that in step 4, right In described characteristic point, gradient image takes centered by characteristic point 24 σ_i×24σ_iWindow, and window is divided into 4 × 4 sub regions, every sub regions size is 9 σ_i×9σ_i, adjacent subregion has width to be 2 σ_iOverlap, Mei Gezi district A σ is all used in territory₁=2.5 σ_I'sGaussian kernel is weighted, and then calculates the subregion description vectors of a length of 4: d_v=(∑ L_x,∑L_y,∑|L_x|,∑|L_y|), then by Gauss window that another size is 4 × 4 vector to every sub regions d_vIt is weighted, is finally normalized, generate the description vectors of 4 × 4 × 4=64 dimension.