Catálogo de publicaciones - libros

This paper discussed the low level machine vision on fruit and vegetable harvesting robot, introduced the recognition and location of fruit and vegetable objects under nature scenes, put forward a new segmentation method combined with several color models. What’s more, it presented a novel conception for the determination of the abscission point, successfully resolved the location of center and abscission point when the fruit were partially occluded. Meanwhile, by the technique of geometry, it settled the locations of the abscission point when the fruit grew askew. It proved good effect under the nature scene.

The problems of the digital information transmission safety, its’ copyright protection etc. are seriously increase. Digital watermark is one of the good solutions for many problems mentioned above. Generally the watermark information is a one-dimension random signal. Two-dimension gray watermark is difficult to guarantee the robustness of the information embedding. Put a new kind of discrete wavelet transform (DWT) and the singular value decompose (SVD) on two-dimension gray watermark, the watermark becomes very strong robust to the compression, noise, low pass filter, high pass filter and median filter. It is useful for the safety of digital information transmission and its copyright protection.

A semiautomatic nonrigid registration is proposed to match preoperative CT images and real-time MR images during the ablation of liver tumors. The method includes three parts. Firstly, the livers are extracted from CT and MR images. Since the CT image is acquired a several days before surgery, there is enough time for surgeons to segment the liver from the CT image manually. During the surgery, active contour is used to extract the liver from the MR image. Then, we use affine registration to match the images roughly. Finally, high accuracy registration is reached by free-form deformation registration. The average distance of liver surfaces is within 1.5 mm and the result of visual inspection is very well. The results are far better than Andres’.

In this paper, a new approach to terrain generation based on terrain examples is proposed. Existing procedural algorithms for generation of terrain have several shortcomings. The most popular approach, fractal-based terrain generation, is efficient, but is difficult for users to control. In this paper, we provide a semiautomatic method of terrain generation that uses a four-process genetic algorithm approach to produce a variety of terrain types using only intuitive user inputs. We allow users to specify a rough sketch of terrain silhouette map, retrieve terrain examples based on support vector machine (SVM) from the terrain dataset, cut a region from the terrain examples and fill in the terrain silhouette map. We also generate a photorealistic texture based on the aerial or satellite images. Consequently, we generate the terrain which has both geometrical data and texture data and provide a balance between user input and real-world data capture unmatched.

This paper presents a new reverse engineering method for creating 3D mesh models, which approximates unorganized noisy point data without orientation information. The main idea of the method is based on statistics Bayesian model. Firstly the feature enhancing prior probabilities over 3D data retrieve the sharp features, such as edges or corners. Then a local polynomial probabilistic model is used to approximate a continuous differentiable manifold. The Bayesian model uses an iterative fitting clustering algorithm to improve the noise tolerance in geometry accuracy. After iteration a density-based estimation function can automatically remove the outliers. Furthermore, the current sphere cover meshing approach is improved to reconstruct the mesh surface from the noisy point data. Experimental results indicate that our approach is robust and efficient. It can be well applied to smoothing noisy data, removing outliers, enhancing features and mesh reconstruction.

In this paper, we describe a novel algorithm for searching and comparing 3D models, called the Integer Medial Axis Skeleton (IMAS), in which the geometric and topological information is encoded in the form of an IMAS and uses skeletal binary tree matching techniques to match the skeletons and to compare them. The skeletal binary tree is constructed based on the IMAS that was constructed using a modification of an IMA transforms algorithm. The similarity calculation between 3D models is processed using a coarse-to-fine strategy. A feature of skeletal binary tree matching is the ability to perform part-matching and provide a fast estimation of similarity between models. The performance of the proposed algorithm is compared to some previous approaches by means of precision/recall tests. Generally, results show that the new algorithm introduces improvements in the 3D-model retrieval process.

Cut-and-paste operation for 3D models is a new geometry modelling method by examples. The method constructs a new model from two or more existed models by cutting some useful parts from source models and pasting to target model. A novel cut-and-paste algorithm for mesh models based on implicit surface Boolean operations is proposed. The algorithm composes of following steps: Firstly, cut part of mesh from source model and align it with the target model. Secondly, convert the two mesh models into point models and convert the point models into implicit surfaces with RBF interpolation. Thirdly, merge the two implicit surfaces by union operation. And finally, polygonize the implicit surface. Then the final new mesh model is obtained. Because the two elements of Boolean operation are not mesh models, it can avoid complex and time consume intersection calculations of facets and edges. The experiments show that good results can be obtained with the algorithm. The approach can be used to create special and exaggerated models for computer animation.

In Virtual Reality, 3D graphics is generally illustrated by meshes of 3D geometric model. This paper provides a novel algorithm to construct progressive meshes based on the reverse subdivision. The dense mesh is simplified into coarse one by decimating the redundant information level by level. Loop subdivision scheme, which is an approximating scheme, is adopted as interpolatory one in the simplification process. The implementation of our algorithm consists of three key steps, splitting, predicting and updating. The progressive mesh can be reconstructed lossless. The experiments have shown that the new algorithm is efficient and more faster than previous related techniques.

In this paper, we propose a new method for surface reconstruction from scattered point set based on least square radial basis function (LSRBF) and orthogonal least square forward selection procedure. Firstly, the traditional RBF formulation is rewritten into least square formula. A implicit surface can be represented with fewer centers. Then, the orthogonal least square procedure is utilized to select significant centers from original point data set. The RBF coefficients can be solved from the triangular matrix from OLS selection through backward substitution method. So, this scheme can offer a fast surface reconstruction tool and can overcome the numerical ill-conditioning of coefficient matrix and over-fitting problem. Some examples are presented to show the effectiveness of our algorithm in 2D and 3D cases.

In this paper, we present a novel method to match and retrieve 3D CAD models. Firstly the CAD model is voxelized. Then the voxelized model is divided into several regions, and the entropy of each region is calculated. All the entropies constitute the feature vector representing the models when compared with other models using Euclidean distance. A system based on this idea is implemented and its feasibility is validated.