Wenguang Hou

Non-rigid multi-modal medical image registration by combining L-BFGS-B with cat swarm optimization

Non-rigid multi-modal image registration plays an important role in medical image processing and analysis. Optimization is a key component of image registration. Mapped as a large-scale optimization problem, non-rigid image registration often requires global optimization methods because the functions defined by similarity metrics are generally non-convex and irregular. In this paper, a novel optimization method is proposed by combining the limited memory Broyden-Fletcher-Goldfarb-Shanno with boundaries (L-BFGS-B) with cat swarm optimization (CSO) for non-rigid multi-modal image registration using the normalized mutual information (NMI) measure and the free-form deformations (FFD) model. The proposed hybrid L-BFGS-B and CSO (HLCSO) method uses cooperative coevolving to tackle non-rigid image registration, and employs block grouping as the grouping strategy to capture the interdependency among variables. Moreover, to achieve faster convergence and higher accuracy of the final solution, the local optimization method L-BFGS-B and the roulette wheel method are introduced into the seeking mode and the tracing mode of the HLCSO, respectively. Extensive experiments on 3D CT, PET, T1, T2 and PD weighted MR images demonstrate that the proposed method outperforms the L-BFGS-B method and the CSO method in terms of registration accuracy, and it is provided with reasonable computational efficiency.

Poisson disk sampling in geodesic metric for DEM simplification

Abstract To generate highly compressed digital elevation models (DEMs) with fine details, the method of Poisson disk sampling in geodesic metric is proposed. The main idea is to uniformly pick points from DEM nodes in geodesic metric, resulting in terrain-adaptive samples in Euclidean metric. This method randomly selects point from mesh nodes and then judges whether this point can be accepted in accordance with the related geodesic distances from the sampled points. The whole process is repeated until no more points can be selected. To further adjust the sampling ratios in different areas, weighted geodesic distance, which is in relation to terrain characteristics, are introduced. In addition to adaptability, sample distributions are well visualised. This method is simple and easy to implement. Cases are provided to illustrate the feasibility and superiority of the proposed method.

Digital deformation model for fisheye image rectification.

Fisheye lens can provide a wide view over 180°. It then has prominence advantages in three dimensional reconstruction and machine vision applications. However, the serious deformation in the image limits fisheye lenss usage. To overcome this obstacle, a new rectification method named DDM (Digital Deformation Model) is developed based on two dimensional perspective transformation. DDM is a type of digital grid representation of the deformation of each pixel on CCD chip which is built by interpolating the difference between the actual image coordinate and pseudo-ideal coordinate of each mark on a control panel. This method obtains the pseudo-ideal coordinate according to two dimensional perspective transformation by setting four marks deformations on image. The main advantages are that this method does not rely on the optical principle of fisheye lens and has relatively less computation. In applications, equivalent pinhole images can be obtained after correcting fisheye lens images using DDM.

Gradient boosting for single image super-resolution

Abstract The learning-based single image super-resolution (SISR) algorithm aims at recovering a high-resolution (HR) image from low-resolution (LR) input. The quality of the HR output mainly depends on the strength of the learning algorithms. Observing that gradient boosting is powerful in dealing with learning problems, we propose a new SISR method based on the gradient boosting framework. First, the boosting framework is extended to the general form of multi-output regression. Then, an error correction approximation is used to sequentially train the boosting trees. The training data for each tree are the pairs of the features of the LR image patches and the negative gradients of the loss function. Meanwhile, shrinkage, a slow learning strategy, is exploited to reduce the risk of overfitting. Finally, all boosting trees are linearly combined to form an accurate predictor. The experimental results verify that our method can generate visually pleasant HR images and achieve accuracy on par with state-of-the-art methods in terms of quantitative evaluation.

Surface Reconstruction through Poisson Disk Sampling

This paper intends to generate the approximate Voronoi diagram in the geodesic metric for some unbiased samples selected from original points. The mesh model of seeds is then constructed on basis of the Voronoi diagram. Rather than constructing the Voronoi diagram for all original points, the proposed strategy is to run around the obstacle that the geodesic distances among neighboring points are sensitive to nearest neighbor definition. It is obvious that the reconstructed model is the level of detail of original points. Hence, our main motivation is to deal with the redundant scattered points. In implementation, Poisson disk sampling is taken to select seeds and helps to produce the Voronoi diagram. Adaptive reconstructions can be achieved by slightly changing the uniform strategy in selecting seeds. Behaviors of this method are investigated and accuracy evaluations are done. Experimental results show the proposed method is reliable and effective.

Minimum spanning tree-based digital terrain model detection from light detection and ranging points

LiDAR is widely used to collect point data-set representing the scanning scenes. Several products can be extracted from raw LiDAR data. DTM is among the most important. For traditional DTM detection approaches such as TIN-based and parametric methods, a negative outlier detector is employed beforehand since negative outliers will drive these methods to converge into an erroneous terrain surface. However, automatic outlier removal remains challenging in related fields. To overcome the obstacle, this article proposes minimum spanning tree-based detection method. An undirected graph is constructed for LiDAR points on basis of delauney rule. Edges are weighted by absolute slope value. Then, greedy algorithm is taken to generate MSF for the undirected graph. Edges with steep slopes are omitted in this process, which is how more than one MST arises. The DTM is generally the MST with largest area. Cases validate the effectiveness.

Adaptive image sampling through establishing 3D geometrical model

Adaptive sampling for high dimensional manifold attracts much attention from related fields. The principal curvature based strategy is one of the popular methods. However, principal curvature estimation remains an open problem. Considering the relationship between geodesics and the principal curvatures of manifold, we transform the optimized sampling density computation into the problem of uniform sampling in the geodesic metric of manifold. Therefore, two well studied uniform sampling methods such as Poisson disk and farthest point strategy are used. For image sampling, a 3D geometrical metric model is built based on mean shift. Mean shift value is applied to describe the image grey information and taken as the height of this model. Uniform sampling is implemented to generate samples with blue noise properties on the 3D model surface. Then, adaptive results are obtained when these samples are projected back to the original 2D image. In contrast to previous methods, this strategy is flexible and can be easily extended to unorganized points simplification or mesh coarsening. Extensive experiments demonstrated the effectiveness of the proposed method.

Prostate Surgery Path Planning Based on Simplified Delay-PCNN

This paper presents an algorithm to find the shortest path in 3D(three-dimensional) prostate surgery planning. Using a simplified delay pulse coupled neural network(S-DPCNN) model, a shortest path can be drawn automatically from the target position to the puncture point. Compared to the traditional pulse coupled neural network(PCNN), S-DPCNN needs much fewer neurons and therefore decreases complexity of computation. Experiments on computer simulations show the validity of this method.