GPU加速的A-NSIFT与PO-GMMREG：无旋转图像配准的高效解决方案

34 浏览量更新于2024-08-27 1 收藏 839KB PDF 举报

本文主要探讨了在图像配准领域的一个重要研究主题，即如何提高图像配准的效率和准确性，特别是在处理姿态变化较大的情况下。文章的焦点在于开发一种快速无旋转特征的图像配准方法，它结合了改进的N-SIFT算法（Accelerated-NSIFT，简称A-NSIFT）和基于Gaussian Mixture Model (GMM) 的并行优化策略（Parallel Optimization for GMMREG）。N-SIFT（尺度不变特征变换）是一种广泛应用于计算机视觉中的关键算法，它对图像特征进行尺度和旋转不变性处理，但对于大型姿态差异下的配准，传统的N-SIFT可能会面临性能瓶颈。 A-NSIFT的改进旨在通过优化计算流程和利用GPU/CUDA并行编程技术来加速特征提取过程。CUDA是NVIDIA开发的一种并行计算平台和编程模型，能够显著提升处理大规模数据时的计算速度。通过这种方式，A-NSIFT能够在保持特征描述符稳定性的同时，极大地缩短了配准的时间消耗。此外，GMMREG部分则引入了概率模型，通过Gaussian Mixture Model（高斯混合模型）来建模图像间的潜在相似度分布。这种方法有助于更精细地估计图像之间的匹配关系，尤其是在存在大量噪声或复杂场景的情况下。通过并行优化，GMMREG能够有效地整合多线程处理，进一步提升整体配准算法的效率。本文的研究目标是解决图像配准中的实时性和精度问题，特别针对那些对时间敏感的应用，如医学影像分析、机器人导航或无人机定位等。通过将A-NSIFT与PO-GMMREG相结合，作者期望能在保持较高准确性的前提下，显著降低配准所需的时间，从而扩展其在实际应用中的适用范围。由于该研究还在接受最终编辑阶段，因此文中可能还包含尚未完全确定的部分，但一旦发表，读者将能从DOI: 10.1109/TBME.2015.2465855获取到完整的、详细的实验结果、理论分析以及对比其他方法的性能评估。这篇文章为图像配准领域的研究者和工程师提供了一种创新且实用的方法，有望推动该领域的技术进步。

http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI

10.1109/TBME.2015.2465855, IEEE Transactions on Biomedical Engineering

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <

The rest of the paper is organized as follows. In Section II,

we present the main idea of our A-NSIFT method for feature

point extraction and our Parallel Optimization based on

Gaussian Mixture Model Registration. In Section III are given

experimental results and discussion, followed by a conclusion

in Section IV.

II. F

AST ROTATION-FREE FEATURE BASED RIGID REGISTRATION

FRAMEWORK

The proposed feature based rigid registration algorithm

consists of three steps. First, the interest point sets from the

fixed and moving images are extracted using our A-NSIFT

method, which transforms the image registration into point set

registration. Then, a Parallel Optimization based on the

Gaussian Mixture Model Registration is proposed in order to

obtain the transformation which best matches the fixed point set

and moving point set. Finally, the moving image is registered

according to the transformation obtained in the second step.

The flowchart of our proposed registration algorithm is shown

in Fig. 1.

A. Accelerated-NSIFT for interest point extraction

The NSIFT method [16], which extends the SIFT method

from 2D to multi-dimensional images, consists of three steps.

First, the doG (difference of Gaussian) scale space is

constituted by a series of Gaussian blurred images which are

generated by a multilevel image pyramid. Then, local feature

points are extracted by localizing the extrema in the

approximation of doG scale space. Third, the features of each

point are depicted by a feature descriptor through summarizing

the gradients near this located feature point.

Algorithm 1 The proposed A-NSIFT method based on CUDA-programming

Input: the moving (fixed) image.

Output: the moving (fixed) set

1 Begin:

2 load the image in CPU and copy to GPU memory as the initial image

(,,)

xyz

Iv v v

3 while (octavenum<OCTAVE)

4 compute the Gaussian

smoothing image series with the initial image

(,,)

xyz

Iv v v

;

(,,,) (,,,)*(,,)

xyz xyz xyz

Lvvv Gvvv Ivvv

σσ

222

(, ,,)

xyz

vvv

xyz

Gv v v e

πσ

−

6 generate the doG space;

(,,,) (, ,, ) (,,,)

xyz xyz xyz

Dvvv Lvvvl Lvvv

σ σσ

= −

8 local the extreme in the doG space, label them in the mask volume and

compute the localization.

9 downsample the last scale Gaussian smoothing image in this octave

and use it as the initial image.

10 octavenum ++;

11 endwhile

12 End

Note: OCTAVE indicates the number of down-sampling scales of an image

[15].

Feature points extracted from the NSIFT method are

invariant to image scale and rotation, whereas it costs too much

time to generate the interest points’ feature descriptor by the

NSIFT method. To reduce the computation cost, we propose an

accelerated NSIFT (A-NSIFT) to speed up the NSIFT while

extracting the distinctive scale invariant features in the 3D

medical image. Our proposed A-NSIFT does not need to

construct the feature descriptor, but uses a multilevel image

pyramid to generate the doG space and to locate the extrema in

the doG space, followed by the acceleration of the first two

steps of the NSIFT using CUDA-programming. Algorithm 1

shows the preudocode of our proposed A-NSIFT method, in

which the CUDA-programming idea can be found from line 4

to line 9. Four kernels are used in the CUDA-programming: the

Gaussian smoothing kernel, the doG generating kernel, the

extrema localization kernel and the downsampling kernel.

(1) Gaussian smoothing kernel. This kernel is applied to the

initial images (both fixed images and moving images) for the

data parallelism by assigning each voxel a thread. Supposing

that voxels are stored in the order v

, v

, and v

, then we can use

M threads per block and N blocks for all the voxels in data

parallelism. Separable filtering and data reuse are integrated

into our parallel scheme to speed up the feature extraction.

Separable filtering is used to avoid the multi-dimensional

convolution and is accomplished by performing convolution of

the initial input image with a 3D Gaussian filter. Since the

Gaussian filter is separable, the convolution is performed in

separate passes for the v

, v

, and v

filters. The second

technique is the data reuse using the texture memory, which is

completed by allocating texture memory and binding the initial

data on it. Since the convolution value at each voxel is

correlated with its neighbor voxels, each voxel can be used

many times. We would note that the data reuse with the texture

memory is also used in the following three steps.

(2) doG generating kernel. This kernel generates the doG

space by taking the series of Gaussian blurred images as input

and producing the difference of Gaussian space at each voxel

respectively in v

, v

, and v

directions.

(3) Extrema localization kernel. This group localizes the

position of the distinct point. It takes the doG space as the input

and label the local extrema in the mask volume.

(4) Downsampling kernel. This kernel downsamples the

Gaussian smoothing images. It takes the last scale of Gaussian

smoothing image as its input and the downsampled image as

the next octave’s initial image.

By using our A-NSIFT, interest point sets from both the

moving image and the fixed image are extracted, and then the

image registration is converted into finding a mapping from the

two interest point sets.

B. Parallel optimization based on GMM registration

1) Point set registration with the Gaussian mixture model

Based on the interest point extraction, the moving point set

can be obtained as well as the fixed set, in which each point can

be treated as a statistical sample from a continuous probability

distribution. Thus, the probability distribution (which is a

continuous density function) of the interest point set can be

represented by a Gaussian mixture model [30], [31] as

剩余11页未读，继续阅读

weixin_38640674

粉丝: 2
资源: 960

GPU加速的A-NSIFT与PO-GMMREG：无旋转图像配准的高效解决方案

西电zelianwen的配准代码(含SAR-SIFT)_SARSIFT_SAR-SIFT_sift配准_包含SAR-SIFT算法

基于卷积神经网络结合改进Harris-SIFT的点云配准方法

基于SIFT特征的图像配准（附Matlab源代码）之仿真图像

sar-sift进行sar图像与光学图像配准

基于sift算法实现图像配准matlab

基于 ga-sift 算法的无人机航拍图像实时拼接

粒子群算法改进SIFT图像配准

图像配准 matlab sift

基于SIFT图像配准matlab

sift算法_图像配准SIFT

最新资源