CN106558073A - Based on characteristics of image and TV L1Non-rigid image registration method - Google Patents

Abstract

The invention discloses a kind of be based on characteristics of image and TV L¹Non-rigid image registration method, belong to image processing field.The problem of registration accuracy and registering time requirement cannot be met simultaneously for traditional non-rigid image registration method, and the present invention is in TV L¹Two kinds of (SIFT TV L are proposed on the basis of method¹With SURF TV L¹) it is from coarse to fine, by the non-rigid image registration method of feature level to Pixel-level.Feature point extraction is carried out to floating image and reference picture initially with SIFT methods or SURF methods, rough registration (feature level registration) is carried out using the characteristic point extracted；Then, on the basis of rough registration, using the TV L based on optical flow field¹Method carries out fine registration (pixel level registration).Show that the present invention can realize accurate, quick, the stable registration of medical image through substantial amounts of registration test, overcome the defect that traditional non-rigid registration method is present.

Description

Based on characteristics of image and TV-L1Non-rigid image registration method

Technical field

The invention belongs to image processing field, and in particular to based on characteristics of image and TV-L¹Non-rigid image registration side Method.

Background technology

Image registration techniques are by different acquisition times (Time), different sensors (Sensor), different acquisition conditions (Condition) mistake matched by Same Scene (Scene) or two width or multiple image of same target (Object) Journey, is widely used in the aspects such as Medical Image Processing, remote sensing image processing.In the registration of medical image, registration is needed Larger nonhomogeneous deformation would generally occur between two width images, therefore non-rigid image registration technology is that Medical Image Processing is ground The hot issue studied carefully.The method of non-rigid image registration is broadly divided into the method for registering based on characteristics of image and based on figure at this stage As the method for registering of gray scale.Based on the registration Algorithm speed of characteristics of image, but for the unconspicuous image of feature has office It is sex-limited；Method for registering based on gradation of image can carry out high-precision registration to little strain image, but the registering time is longer.

Based on the registration of characteristics of image, it is the searching registration parameter to extract the feature of reflection image important information as foundation, Make similarity measure maximum.Had based on the classical way of characteristics of image at this stage：SIFT(Scale-invariant Feature Transform) method and SURF (Speeded-Up Robust Features) method.

Based on the method for registering of gradation of image, the pixel or voxel constitutive characteristic space first with image, then basis The statistical information of gray value calculates similarity measure, tries to achieve registration parameter, realizes image registration.Based on optical flow field in gray scale method Modelling is the focus of Recent study, mainly using optical flow estimation, by the meter to each pixel instantaneous velocity vector information Calculate, estimate deformation field, realize the registration of image.1998, Thirion proposed famous Demons side according to optical flow estimation Method, Wang He proposed Active Demons (AD) method in 2005, solved Demons methods to a certain extent to big The unworthiness and height of deformation is time consuming nature, introduces Demons (MD) algorithm of multiresolution strategy.Other scholars are to Demons Method has carried out different improvement so as to adapt to the registration of different type image.2007, Tomas Pock et al. were in traditional light On the basis of flow field model, using total variation regular terms (Total Variation regularization, TV regular terms) and Shandong nation property data item (L1 norm items) builds energy function, it is proposed that a kind of new TV-L¹Method, registration effect are more preferable.

The content of the invention

The present invention cannot meet registration accuracy with registering time requirement simultaneously for traditional non-rigid image registration method Problem, there is provided a kind of to be based on characteristics of image and TV-L¹Non-rigid image registration method.

The present invention is achieved by the following technical solutions, and the present invention is comprised the following steps：

The first step, carries out feature point extraction to floating image and reference picture using SIFT methods or SURF methods, utilizes The characteristic point of extraction carries out rough registration (feature level registration).

Second step, on the basis of rough registration, using the TV-L based on optical flow field¹Method carries out fine registration (Pixel-level Registration).

In the first step, SIFT methods or SURF methods adopt difference of Gaussian function (Difference of Gaussians, DoG) multiple dimensioned feature extraction is carried out to image.

TV-L in the second step¹Method is carried out from coarse to fine on multiple yardsticks using multi-resolution pyramid to image Registration.

The first step is comprised the following steps that：

Step 1-1：Reference picture S and floating image M is read in, and pretreatment is carried out to S and M；

Step 1-2：Metric space is built, the parameter of difference of Gaussian is set；Yardstick is built using Hessian matrixes empty Between；

Step 1-3：Extreme point is screened, and generates SIFT or SURF description of these extreme points；

Step 1-4：Feature Points Matching is carried out using description that step 3 is tried to achieve, deformation field is generated, floating image M is entered Row rough registration, tries to achieve rough registration image M '.

The second step is comprised the following steps that：

According to rough registration image M that step 1-4 is tried to achieve ' and reference picture S adopt TV-L¹Algorithm carries out smart registration tries to achieve and matches somebody with somebody Quasi- image I：

For Level=1:1:L (L is default maximum decomposition scale)

For n=1:1:N (N is default maximum iteration time)

Using formula

u_d=v_d-θdivp

Wherein, u, v are optimized object function, and θ is the parameter of a value very little, and λ is a constant,Calculate for gradient Son, τ are constant and τ≤2 (N+1), and to solve u, the intermediate variable of v, k are iterationses to p, and ρ is residual image.

To u, v is updated, and carries out spatial alternation to floating image M

Calculate similarity measure：If meeting condition, break；If being unsatisfactory for continuing iteration；

end

end。

The present invention is from coarse to fine, by the non-rigid image registration method (SIFT-TV-L of feature level to Pixel-level¹With SURF-TV-L¹).Medical image is carried out high accuracy to be met simultaneously, the requirement of high speed, present invention employs by slightly to Carefully, by the registration strategies of feature level to Pixel-level.Initially with SIFT methods or SURF methods to floating image and reference picture Feature point extraction is carried out, and rough registration (feature level registration) is carried out using the characteristic point extracted；Then, on the basis of rough registration, Using the TV-L based on optical flow field¹Method carries out fine registration (pixel level registration).In order to further optimize registration result, adopt Multiresolution strategy.First step SIFT method or SURF methods adopt difference of Gaussian function (Difference of Gaussians, DoG) multiple dimensioned feature extraction is carried out to image；Second step TV-L¹Method is existed using multi-resolution pyramid Registration from coarse to fine is carried out to image on multiple yardsticks.

Description of the drawings

Fig. 1, Fig. 2, Fig. 3, Fig. 4 are brain MR image registration result and difference；

Fig. 5, Fig. 6, Fig. 7, Fig. 8 are liver MR image registration results and difference；

Specific embodiment

Below in conjunction with the accompanying drawings and specific embodiment the present invention is further detailed explanation.

The present invention realizes that the registration process of medical image sequentially passes through following steps：

Step 1：Reference picture S and floating image M is read in, and pretreatment is carried out to S and M；

Step 2：Metric space is built, the parameter of difference of Gaussian is set；Yardstick is built using Hessian matrixes empty Between；

Step 3：Extreme point is screened, and generates SIFT or SURF description of these extreme points；

Step 4：Feature Points Matching is carried out using description that step 3 is tried to achieve, deformation field is generated, floating image M is carried out Rough registration；

Step 5：According to rough registration image M that step 4 is tried to achieve ' and reference picture S adopt TV-L¹Algorithm carries out smart registration and asks Obtain registering image I：

For Level=1:1:L (L is default maximum decomposition scale)

For n=1:1:N (N is default maximum iteration time)

Using formula

u_d=v_d-θdivp

Wherein, u, v are optimized object function, and θ is the parameter of a value very little, and λ is a constant,Calculate for gradient Son, τ are constant and τ≤2 (N+1), and to solve u, the intermediate variable of v, k are iterationses to p, and ρ is residual image.

To u, v is updated, and carries out spatial alternation to floating image M

Calculate similarity measure：If meeting condition, break；If being unsatisfactory for continuing iteration；

end

end；

Step 6：Objective evaluation index, the quality of analysis method are calculated using reference picture S and registering image I.

The step 2 introduces multiresolution strategy into step 6 and realizes registration, so can further improve the essence of registration Degree and convergence rate.

Interpretation

1. the degree of accuracy of image registration

Registration of the realization that can be relatively accurate of the invention to medical image.Fig. 1 to Fig. 8 is to two different medical images Registering experimental result, Fig. 1-Fig. 4 is brain MR image registration result and difference, and Fig. 5-Fig. 8 is liver's MR image registration results And difference.The present invention has preferable registration effect, Tables 1 and 2 difference to large deformation medical image and little deformation medical image It is the present invention in Y-PSNR (PSNR, Peak Signal to Noise Ratio), mean square deviation (MSE, Mean Square Error), correlation coefficient (CC, Correlation Coefficient), mutual information (MI, Mutual Information) and knot Comment with several traditional the objective of non-rigid registration method in 5 indexs of structure similarity (SSIM, Structural SIMilarity) Valency compares.

1 brain MR image of table registration objective evaluation

2 liver CT image registration objective evaluations of table

2. time of image registration

The image registration time directly affects the practical application effect of the present invention.The present invention is relative to tradition based on gray scale Method for registering, registering time are greatly shortened.Table 3 compared for the time with several traditional non-rigid registration methods of the invention.

The 3 registering time of table compares

In sum, the present invention devises two kinds of non-rigid registration method (SIFT-TV-L for being applied to medical image¹With SURF-TV-L¹).The method is based on characteristics of image and optical flow field, can realize the registration to medical image well.

Finally it should be noted that above embodiment is only unrestricted to illustrate technical scheme, although ginseng The present invention is described in detail according to preferably embodiment, but protection scope of the present invention has been not limited thereto, it is any Those familiar with the art the invention discloses technical scope in, the modification that can be readily occurred in or equivalent, Without deviating from the spirit and scope of technical solution of the present invention, should all be included within the scope of the present invention.

Claims (3)

1. a kind of non-rigid image registration method based on characteristics of image and TV-L1, it is characterised in that comprise the following steps：

The first step, carries out feature point extraction to floating image and reference picture using SIFT methods or SURF methods, using extraction Characteristic point carry out rough registration；

Second step, on the basis of rough registration, using the TV-L based on optical flow field¹Method carries out fine registration；

In the first step, SIFT methods or SURF methods carry out multiple dimensioned feature using difference of Gaussian function pair image and carry Take；

TV-L in the second step¹Method is carried out from coarse to fine matching somebody with somebody to image on multiple yardsticks using multi-resolution pyramid It is accurate.

2. the non-rigid image registration method based on characteristics of image and TV-L1 according to claim 1, it is characterised in that The first step is comprised the following steps that：

Step 1-1：Reference picture S and floating image M is read in, and pretreatment is carried out to S and M；

Step 1-2：Metric space is built, the parameter of difference of Gaussian is set；Metric space is built using Hessian matrixes；

Step 1-3：Extreme point is screened, and generates SIFT or SURF description of these extreme points；

Step 1-4：Feature Points Matching is carried out using description that step 3 is tried to achieve, deformation field is generated, floating image M is carried out slightly Registration, tries to achieve rough registration image M '.

3. the non-rigid image registration method based on characteristics of image and TV-L1 according to claim 2, it is characterised in that The second step is comprised the following steps that：

According to rough registration image M that step 1-4 is tried to achieve ' and reference picture S adopt TV-L¹Algorithm carries out smart registration and tries to achieve registering figure As I：

For Level=1:1:L (L is default maximum decomposition scale)

For n=1:1:N (N is default maximum iteration time)

Using formula

u_d=v_d-θdivp

$p^{k+1}=\frac{p^k+\mathrm{\τ}\▿({\mathrm{divp}}^k-v_d/\mathrm{\θ})}{1+\mathrm{\τ}|\▿({\mathrm{divp}}^k-v_d/\mathrm{\θ})|}$

Wherein, u, v are optimized object function, and θ is the parameter of a value very little, and λ is a constant, and ▽ is gradient operator, τ For constant and τ≤2 (N+1), to solve u, the intermediate variable of v, k are iterationses to p, and ρ is residual image；

To u, v is updated, and carries out spatial alternation to floating image M

Calculate similarity measure：If meeting condition, break；If being unsatisfactory for continuing iteration；

end

end。