Model-based mapping of reconfigurable image registration on FPGA platforms

Image registration is a computationally intensive application in the medical imaging domain that places stringent requirements on performance and memory management efficiency. This paper develops techniques for mapping rigid image registration applications onto configurable hardware under real-time performance constraints. Building on the framework of homogeneous parameterized dataflow, which provides an effective formal model of design and analysis of hardware and software for signal processing applications, we develop novel methods for representing and exploring the hardware design space when mapping image registration algorithms onto configurable hardware. Our techniques result in an efficient framework for trading off performance and configurable hardware resource usage based on the constraints of a given application. Based on trends that we have observed when applying these techniques, we also present a novel architecture that enables dynamically-reconfigurable image registration. This proposed architecture has the ability to tune its parallel processing structure adaptively based on relevant characteristics of the input images.

Video registration using dynamic textures

We consider the problem of spatially and temporally registering multiple video sequences of dynamical scenes which contain, but are not limited to, nonrigid objects such as fireworks, flags fluttering in the wind, etc., taken from different vantage points. This problem is extremely challenging due to the presence of complex variations in the appearance of such dynamic scenes. In this paper, we propose a simple algorithm for matching such complex scenes. Our algorithm does not require the cameras to be synchronized, and is not based on frame-by-frame or volume-by-volume registration. Instead, we model each video as the output of a linear dynamical system and transform the task of registering the video sequences to that of registering the parameters of the corresponding dynamical models. As these parameters are not uniquely defined, one cannot directly compare them to perform registration. We resolve these ambiguities by jointly identifying the parameters from multiple video sequences, and converting the identified parameters to a canonical form. This reduces the video registration problem to a multiple image registration problem, which can be efficiently solved using existing image matching techniques. We test our algorithm on a wide variety of challenging video sequences and show that it matches the performance of significantly more computationally expensive existing methods.     

Image registration using hierarchical B-splines

Hierarchical B-splines have been widely used for shape modeling since their discovery by Forsey and Bartels. We present an application of this concept, in the form of free-form deformation, to image registration by matching two images at increasing levels of detail. Results using MRI brain data are presented that demonstrate high degrees of matching while unnecessary distortions are avoided. We compare our results with the nonlinear ICP (iterative closest point) algorithm (used for landmark-based registration) and optical flow (used for intensity-based registration).     

Non-rigid registration using distance functions

This paper deals with the registration of geometric shapes. Our primary contribution is the use of a simple and robust shape representation (distance functions) for global-to-local alignment. We propose a rigid-invariant variational framework that can deal as well with local non-rigid transformations. To this end, the registration map consists of a linear motion model and a local deformations field, incrementally recovered. In order to demonstrate the performance of the selected representation a simple criterion is considered, the sum of square differences. Empirical validation and promising results were obtained on examples that exhibit large global motion as well as important local deformations and arbitrary topological changes.     

ICP registration using invariant features

Investigates the use of Euclidean invariant features in a generalization of iterative closest point (ICP) registration of range images. Pointwise correspondences are chosen as the closest point with respect to a weighted linear combination of positional and feature distances. It is shown that, under ideal noise-free conditions, correspondences formed using this distance function are correct more often than correspondences formed using the positional distance alone. In addition, monotonic convergence to at least a local minimum is shown to hold for this method. When noise is present, a method that automatically sets the optimal relative contribution of features and positions is described. This method trades off the error in feature values due to noise against the error in positions due to misalignment. Experimental results suggest that using invariant features decreases the probability of being trapped in a local minimum and may be an effective solution for difficult range image registration problems where the scene is very small compared to the model     

Digital image registration using projections

In many application fields (e.g., aerospace and biomedical image processing), one has to deal with a sequence of images whose observation is made difficult by unpredictable relative movements of the camera and the scene. In order to obtain a stable display, the displacement of each image in the sequence with respect to one chosen as a reference must be preliminarily determined. This would usually require two-dimensional algorithms, involving a considerable computational effort. This work discusses a new algorithm for image registration, which requires only one-dimensional Fourier transformations. Preliminary experimental results are reported.     

Three-dimensional object registration using wavelet features

Recent developments in shape-based modeling and data acquisition have brought three-dimensional models to the forefront of computer graphics and visualization research. New data acquisition methods are producing large numbers of models in a variety of fields. Three-dimensional registration (alignment) is key to the useful application of such models in areas from automated surface inspection to cancer detection and surgery. The algorithms developed in this research accomplish automatic registration of three-dimensional voxelized models. We employ features in a wavelet transform domain to accomplish registration. The features are extracted in a multi-resolutional format, thus delineating features at various scales for robust and rapid matching. Registration is achieved by using a voting scheme to select peaks in sets of rotation quaternions, then separately identifying translation. The method is robust to occlusion, clutter, and noise. The efficacy of the algorithm is demonstrated through examples from solid modeling and medical imaging applications.     

Real-time 3D registration using GPU

3D registration is a computer vision technique of aligning multi-view range images with respect to a reference coordinate system. Aligning range images is an important but time-consuming task for complete 3D reconstruction. In this paper, we propose a real-time 3D registration technique by employing the computing power of graphic processing unit (GPU). A point-to-plane 3D registration technique is completely implemented using CUDA, the up-to-date GPU programming technique. Using a hand-held stereo-vision sensor, we apply the proposed technique to real-time 3D scanning of real objects. Registration of a pair of range images, whose resolution is 320 × 240, takes about 60 ms. 3D scanning results and processing time analysis are shown in experiments. To compare the proposed GPU-based 3D registration with other CPU-based techniques, 3D models of a reference object are reconstructed. Reconstruction results of three different techniques in eight different scanning speed are evaluated.     

Deformable registration using patch-wise shape matching

We present a novel approach for non-rigid registration of partially overlapping surfaces acquired from a deforming object. To allow for large and general deformations our method employs a nonlinear physics-inspired deformation model, which has been designed with a particular focus on robustness and performance. We discretize the surface into a set of overlapping patches, for each of which an optimal rigid motion is found and interpolated faithfully using dual quaternion blending. Using this discretization we can formulate the two components of our objective function—a fitting and a regularization term—as a combined global shape matching problem, which can be solved through a very robust numerical approach. Interleaving the optimization with successive patch refinement results in an efficient hierarchical coarse-to-fine optimization. Compared to other approaches our as-rigid-as-possible deformation model is faster, causes less distortion, and gives more accurate fitting results.     

Moving target tracking using symbolic registration

This correspondence describes a moving target tracking (MTT) algorithm that performs image registration and motion analysis between pairs of images from a passive sensor. Unlike previously reported moving target indicators that operate at the signal level, the registration and motion analysis in the MTT is totally performed at a symbolic level. The operation of the MTT is demonstrated by simulation results obtained from applications of the algorithm to infrared images.     

基于参数化的三维颅面配准

针对颅面配准问题,提出通过对颅面进行参数化将其转换成二维参数域的对应问题。首先,根据人类的生理特征标定6个特征点,利用这些特征点将颅面转换到一个统一的坐标系以实现姿态和大小的统一;其次,以两个外眼角为约束对参考颅面进行最小二乘保角映射,计算出6个特征点的参数值;然后,以这六个生理特征点的参数值为约束,利用最小二乘保角映射将任一待配准模型映射到二维参数域;最后,根据二维参数域确定三维颅面上的对应点,从而实现三维数据配准。为了验证所提方法,以对应点为控制点,利用薄板样条(TPS)变换把参考颅面变形到目标颅面,以变形后两个模型上对应点之间的几何距离的平均为度量,将所提算法和基于主轴分析的迭代最近点(ICP)配准以及基于随机采样控制点的迭代TPS配准方法进行了比较,实验结果表明,所提算法的配准效果优于其他两种方法。     

Spectral registration of noisy sonar data for underwater 3D mapping

3D mapping is very challenging in the underwater domain, especially due to the lack of high resolution, low noise sensors. A new spectral registration method is presented that can determine the spatial 6 DOF transformation between pairs of very noisy 3D scans with only partial overlap. The approach is hence suited to cope with sonar as the predominant underwater sensor. The spectral registration method is based on Phase Only Matched Filtering (POMF) on non-trivially resampled spectra of the 3D data.     

Surface registration using a dynamic genetic algorithm

Robust and fast free-form surface registration is a useful technique in various areas such as object recognition and 3D model reconstruction for animation. Notably, an object model can be constructed, in principle, by surface registration and integration of range images of the target object from different views. In this paper, we propose to formulate the surface registration problem as a high dimensional optimization problem, which can be solved by a genetic algorithm (GA) (Genetic Algorithms in Search Optimization and Machine Learning, Addison-Wesley, Reading, MA, 1989). The performance of the GA for surface registration is highly dependent on its speed in evaluating the fitness function. A novel GA with a new fitness function and a new genetic operator is proposed. It can compute an optimal registration 1000 times faster than a conventional GA. The accuracy, speed and the robustness of the proposed method are verified by a number of real experiments.     

改进遗传算法的医学图像配准

采用遗传算法的医学图像配准时,遗传算法存在收敛速度慢,易早熟的问题,有可能导致误配.提出改进遗传算法(IGA),该方法将外推搜索和黄金分割搜索与标准遗传算法(SGA)相结合,既提高了遗传算法的收敛速度,又有效地防止了早熟.实验结果表明,改进算法具有更好的有效性和精确性.     

Efficient multi-modal image registration using local-frequency maps

Fusing of multi-modal data involves automatically estimating the coordinate transformation required to align the multi-modal image data sets. Most existing methods in literature are not fast enough for practical use (taking more than 30 min to 1 h for estimating non-rigid deformations). We propose a very fast algorithm based on matching local-frequency image representations, which naturally allows for processing the data at different scales or resolutions, a very desirable property from a computational efficiency view point. For the rigid motion case, this algorithm involves minimizing – over all rigid transformations – the expectation of the squared difference between the local-frequency representations of the source and target images. In the non-rigid deformations case, we propose to approximate the non-rigid motion by piece-wise rigid motions and use a novel and fast PDE-based morphing technique that estimates this non-rigid alignment. We present implementation results for synthesized and real (rigid) misalignments between CT and MR brain scans. In both the cases, we validate our results against ground truth registrations which for the former case are known and for the latter are obtained from manual registration performed by an expert. Currently, these manual registrations are used in daily clinical practice. Finally, we present examples of non-rigid registration between T1-weighted MR and T2-weighted MR brain images wherein validation is only qualitatively achieved. Our algorithm's performance is comparable to the results obtained from algorithms based on mutual information in the context of accuracy of estimated rigid transforms but is much faster in computational speed. Accepted: 13 November 2001     

基于高维互信息量的图像配准

基于互信息量的图像配准算法只考虑单个相对应点间的关系,忽略了图像的空间信息,因此当图像分辨率较低、有噪声影响和部分缺损时就容易出现误配.将图像的空间信息引入到配准中,考虑对应点及其邻域的关系,提出了一种新的相似性测度--高维互信息量.新的测度不仅能够反映图像的灰度统计信息,而且能够反映图像的空间信息.实验结果表明,在图像空间分辨率较低、有噪声影响和图像部分缺损的情况下,该算法具有较高的准确性.     

Non-rigid registration using gradient of self-similarity response

Locally affine transformation with globally elastic interpolation is a common strategy for non-rigid registration. Current techniques improve the registration accuracy by only processing the sub-images that contain well-defined structures quantified by Moran's spatial correlation. As an indicator, Moran's metric successfully excludes noisy structures that result in misleading global optimum in terms of similarity. However, some well-defined structures with intensity only varying in one direction may also cause mis-registration. In this paper, we propose a new metric based on the response of a similarity function to quantify the ability of being correctly registered for each sub-image. Using receiver operating characteristic analysis, we show that the proposed metric more accurately reflects such ability than Moran's metric. Incorporating the proposed metric into a hierarchical non-rigid registration scheme, we show that registration accuracy is improved relative to Moran's metric.     

Nonlinear registration using variational principle for mutual information

We extend the multimodal image registration method described in Alexander and Summers [Fast registration algorithm using a variational principle for mutual information, Proc. SPIE Int. Soc. Opt. Eng. 5032 (2003) 1053–1063] to nonlinear registration. A variational principle maximizing mutual information leads to an Euler–Lagrange (EL) equation for the displacement field, represented here in a basis of cubic B-spline functions. A cost function is constructed from the sum of squares of the residuals of the EL equation at a subset of pixels where the magnitude of the spatial gradient of intensity exceeds a user-chosen threshold. The unknown coefficients in the displacement field representation are evaluated using a Levenberg–Marquardt minimization procedure. The proposed method was successfully applied to several image pairs of the same and different modalities, and an artificially constructed series of images containing nonlinear distortions and noise.     

Fast cross-spectral image registration using new robust correlation

In this paper, we explore a new correlation technique for cross-spectral image registration. The proposed technique matches the orientation feature of the second derivatives while making use of a statistical robust M estimator. Furthermore, it takes advantage of Fourier and multi-resolution techniques to reduce the complexity of spatial correlation. Simulation results show that our proposed approach gives more accurate results than the mutual information, and the normalized cross-correlation with prefiltering in terms of speed and accuracy.