Vol. 14 (2005): Abstracts of Papers
Zhang Y., Sim T., Tan C.L.:
Generating personalized anatomy-based 3D facial models from scanned data.
MGV vol. 14, no. 1, 2005, pp. 3-28.
This paper presents a new method for reconstructing animatable, anatomy-based human facial models from scanned range data. Our method adapts a prototype model that is suitable for physically-based animation to the geometry of a specific person's face with minimal user intervention. The prototype model has a known topology and incorporates a multi-layer structure of the skin, muscles, and skull. Based on a series of measurements between a subset of anthropometric landmarks specified on the prototype model and the scanned surface, an automated global alignment adapts the size, position, and orientation of the prototype model to align it with the scanned surface. In the skin layer adaptation, the generic skin mesh is represented as a dynamic deformable model which is subjected to internal force stemming from the elastic properties of the surface and external forces generated by the scanned data points and features. We automatically deform the underlying muscle layer consisting of three types of muscle models. A set of automatically generated skull feature points is then transformed based on the deformed external skin and muscle layers. The new positions of these feature points are used to drive volume morphing applied to the skull template for skull fitting. With the adapted multi-layer anatomical structure, the reconstructed model not only resembles the shape of the individual's face but can also be animated instantly using the muscle and jaw parameters.
Key words: face reconstruction, facial animation, anatomy-based model, multi-layer skin/muscle/skull structure, scanned data, deformable model.
El-Bakry H.M., Zhao Q.:
Speeding-up normalized neural networks for face/object detection.
MGV vol. 14, no. 1, 2005, pp. 29-59.
Finding an object or a face in an input image is a search problem in the spatial domain. Neural networks have shown good results in detecting a certain face/object in a given image. In this paper, faster neural networks for face/object detection are presented. Such networks are designed based on cross correlation in the frequency domain between the input image and the input weights of neural networks. This approach is developed to reduce the computation steps required by these faster neural networks for the search process. The principle of divide and conquer strategy is applied through image decomposition. Each image is divided into small-size sub-images, and then each of them is tested separately using a single faster neural network. Furthermore, the fastest face/object detection is achieved using parallel processing techniques to test the resulting sub-images simultaneously using the same number of faster neural networks. In contrast to using faster neural networks only, the speed-up ratio is increased with the size of the input image when using faster neural networks and image decomposition. Moreover, the problem of local subimage normalization in the frequency domain is solved. The effect of image normalization on the speed-up ratio for face/object detection is discussed. Simulation results show that local subimage normalization through weight normalization is faster than subimage normalization in the spatial domain. The overall speed-up ratio of the detection process is increased as the normalization of weights is carried out off line.
Key words: fast face/object detection, neural networks, cross correlation, image normalization, parallel processing.
Zaqout I., Zainuddin R., Baba S.:
Pixel-based skin color detection technique.
MGV vol. 14, no. 1, 2005, pp. 61-70.
One of the simplest features used for the human face detection problem is the skin color information. A simple and relatively efficient histogram-based algorithm to segment skin pixels from a complex background is presented. The histogram-based algorithm used here is referred to as the lookup table (LUT) and is adopted to identify those intervals which may fall in the skin locus plane. For that purpose, a total of 306,401 skin samples are manually collected from RGB color images to calculate three lookup tables based on the relationship between each single pair of the three components (R, G, B). To estimate the skin locus boundary, a skin classifier box is created by integration of the proposed three heuristic rules based on how often each RGB pixel-relationship falls into its interval.
Key words: skin segmentation, histogram-based approach, lookup table, heuristic rules, face detection.
Ogiela M.R., Tadeusiewicz R.:
Picture languages in machine understanding of medical visualization.
MGV vol. 14, no. 1, 2005, pp. 71-82.
This paper presents theoretical fundamentals and application of context-free and graph languages for cognitive analysis of selected medical visualization. It shows new opportunities for applying these methods of automatic understanding of semantic contents of images in intelligent medical information systems. A successful extraction of the crucial semantic content of medical image may contribute considerably to the creation of new intelligent cognitive systems, or medical computer vision systems. Thanks to the new idea of cognitive resonance between a stream of the data extracted from the image using linguistic methods, and expectations following from the language representation of the medical knowledge, it is possible to understand the subject-oriented content of the visual data. This article shows that structural techniques of soft-computing may be applied in automatic classification and machine perception based on semantic pattern content in order to determine the semantic meaning of the patterns.
Key words: structural pattern recognition, image understanding, artificial intelligence, computer-aided diagnosis, feature extraction, shape analysis.
Abdel-Qader I.M., Maddix M.E.:
Edge detection: wavelets versus conventional methods on DSP processors.
MGV vol. 14, no. 1, 2005, pp. 83-101.
Edge detection is a cornerstone in any computer, robotic or machine vision system. Real time edge detection is a pre-process to many critical applications, such as assembly line inspection and surveillance. Wavelets-based algorithms are replacing traditional algorithms, especially the Haar wavelet because of its simplicity. The Haar algorithm uses a multilevel decomposition to produce image edges corresponding to high frequency wavelet coefficients. In this paper, a real time edge detection algorithm based on Haar is analyzed and compared to conventional edge detectors. Other implemented and compared algorithms are the traditional Prewitt algorithm, and, from a newer generation, the Canny algorithm. The real time implementation of all algorithms is accomplished using TI TMS320C6711 card. In case of Haar, the multilevel decomposition improves the results obtained with noisy images. The results show that the Haar-based edge detector has a low execution time with accurate edge results, and thus represents a suitable algorithm for on-line vision system applications. Canny has produced the thinnest edges, but is not suitable for real time processing using the 6711, and falls short in edge results compared to the Haar results. The wavelet-based algorithm has outperformed other edge detectors.
Key words: Haar Wavelets, edge detection, DSP processor.
Amplitude elimination for stereo image matching based on the wavelet approach.
MGV vol. 14, no. 1, 2005, pp. 103-120.
Point-to-point correspondence is one of the most challenging problems in stereo image matching. Correspondence or disparity established between points of two images is the result of stereo matching. The paper presents new point-to-point correspondence algorithms based on wavelet analysis. Each image in the pair is decomposed into an approximation, and details go through the coarse to fine level. For the above decomposition, multiresolution analysis is used. In the proposed approach, a disparity is found in the wavelet transform space. An extension and generalization of phase-based method is presented. The classical Gabor's approach is extended to real wavelets. Differences of amplitudes (grey levels) in images which frequently appear in stereo pairs are eliminated. Invariance of the disparity determination with respect to amplitude changes may be achieved by choosing an appropriate pair of wavelet systems. The achieved result is broader than the classical one, based on the Gabor wavelet and the phase method. Numerical experiments with images have confirmed this approach. Finally, three concepts (see Section 5) are presented to analyse the problem of disparity determination globally.
Key words: stereo matching, point-to-point correspondence, wavelet transform, amplitude elimination, disparity determination.
Foufou S., Garnier L.:
Obtaining implicit equations of supercyclides and definition of elliptic supercyclides.
MGV vol. 14, no. 2, 2005, pp. 123-144.
The use of Dupin cyclides and supercyclides in CAGD applications has been the subject of many publications in the last decade. Dupin cyclides are low degree algebraic surfaces having both parametric and implicit representations. In this paper, we aim to give the necessary expansions to derive implicit equations of supercyclides in the affine as well as in the projective space, starting from equations of the Dupin cyclide and the transformation matrix. We introduce a particular subfamily of supercyclides, called elliptic supercyclides, and show how to use them for the blending of elliptic quadratic primitives. We also show how one can convert an elliptic supercyclide into a set of rational biquadratic Bézier patches.
Key words: Dupin cyclides, supercyclides, projective and affine geometry.
Cheng B., Wang Y., Zheng N., Bian Z.:
Object based segmentation of video using variational level sets.
MGV vol. 14, no. 2, 2005, pp. 145-157.
The paper demonstrates a new approach to video segmentation which retains some of the attractive features of existing methods and overcomes some of their limitations. The video sequence is represented as a spatio-temporal volume, and is segmented by an extension of active contour model based on Mumford-Shah techniques. The energy function minimization is similar to 3D interface evolution with curvature-dependent speeds. The spatio-temporal volume need not to be smoothed before processing because our method is not sensitive to noise. Each object needs a closed interface, which is embedded as a level set of a higher-dimensional functions, and is propagated by solving a partial differential equation. The interface stops in the vicinity of object boundaries, which are not necessarily defined by the gradient and can be represented with complex topologies. Finally, an experiment is given to show the effectiveness and robustness of the method.
Key words: video sequence, segmentation, Level set, Mumford-Shah functional.
Barsi A., Szirmay-Kalos L., Szécsi L.:
Image-based illumination on the GPU.
MGV vol. 14, no. 2, 2005, pp. 159-169.
In many computer graphics applications it is desirable to augment virtual objects with high dynamic range images representing the real environment. In order to provide an illusion that virtual objects are part of the real scene, illumination of the environment should be taken into account when rendering them. Proper calculation of the illumination from an environment map may be extremely expensive if we wish to account for occlusion, self shadowing and specular materials. In this paper we present a method that performs all the calculations on a per-frame basis, and is already real-time on non-cutting-edge hardware.
Key words: HDRI, shadow, GPU, reflection.
Zaremba M.B., Palenichka R.M., Missaoui R.:
Multi-scale morphological modeling of a class of structural texture.
MGV vol. 14, no. 2, 2005, pp. 171-199.
Consistent and time-efficient modeling of textures is important both for realistic texture mapping in computer graphics and correct texture segmentation in computer vision. A large class of natural and artificial images is represented by the so-called structural textures, which contain visibly repetitive patterns. The multi-scale morphological modeling approach proposed in this paper explicitly describes shape and intensity parameters of structural textures. It is based on a cellular growth of a texture region by a sequential morphological generation of structural texture cells starting from a seed cell. Its main advantage is a concise shape representation for structural texture cells in the form of piecewise linear skeletons. Another advantage is a robust and computationally efficient estimation of texture parameters. The cell parameter estimation is based on the cell localization and adaptive segmentation using a multi-scale matched filter. The experiments reported in the paper are related to texture parameter estimation from synthetic and real textures as well as structural texture synthesis based on the estimated parameters.
Key words: texture modeling, structural texture, local scale, mathematical morphology, parameter estimation, binarization.
Abd Allah M.M.:
A novel approach for fingerprint classification system based on new feature area search.
MGV vol. 14, no. 2, 2005, pp. 201-212.
The paper presents a new fast fingerprint classification method based on direction patterns. The method is designed to be applicable to today's embedded systems for fingerprint authentication, in which small area sensors are employed (large enough to capture all the core and delta points of a fingerprint). The proposed procedure consists of four steps. First, ridge direction is determined at the pixel level. Second, average orientation field flow is assessed within 8x8 blocks. Then pattern matching is applied to determine presence of either of three "feature areas". Finally, the target classes are identified through a novel classification approach, called generally a pattern area. We prove that the search of direction pattern in a specific area is able to classify fingerprints clearly and quickly. With our algorithm, the classification accuracy of 94% is achieved over 4000 images in the NIST-4 database, slightly lower than the conventional approaches. However, the classification speed has improved tremendously, up to about 10 times faster than the conventional singular point approaches at the pixel level.
Key words: fingerprint classification, direction pattern, embedded system.
Cowell J., Hussain F.:
Two template matching approaches to arabic, amharic and latin isolated characters recognition.
MGV vol. 14, no. 1, 2005, pp. 213-232.
With the establishment of commercial OCR systems for Latin text, recent research efforts have been directed at the design of recognition systems for non-Latin scripts, such as Japanese, Cyrillic, Chinese, Hindi, Tibetan, and in particular Arabic. The Unicode 4.0 standard supports 50 scripts that are used across the world, and many, such as Amharic (Ethiopic), have attracted virtually no attention from researchers. An extensive literature review reveals no papers which report on an OCR system for Amharic. This paper describes a normalised technique which can be used for recognition of isolated Arabic, Amharic and Latin characters. Two approaches are considered for identifying the characters by comparing them to a series of templates and using a signature template scheme. The degrees of similarity between pairs of Amharic, Arabic and typical Latin characters are presented in the confusion matrix, and the performance of the two approaches is compared for each of these three character sets.
Key words: OCR, optical character recognition, confusion matrix, fonts, script, Arabic, Amharic, Unicode, template matching.
El-Khamy S.E., Hadhoud M.M., Dessouky M.I., Salam B.M., Abd El-Samie F.E.:
An adaptive cubic convolution image interpolation approach.
MGV vol. 14, no. 3, 2005, pp. 235-258.
Key's (bicubic) image interpolation is one of the well-known, state of the art image interpolation algorithms. In this paper, we introduce an adaptive version of Key's interpolation algorithm. The suggested adaptive algorithm is based on minimization of the squared estimation error at each pixel in the interpolated image. Thus, the overall mean square error (MSE) in the entire image is minimized. The suggested algorithm takes into consideration the low resolution (LR) image degradation model. The Key's formula comprises two controling parameters. A study of the effect of optimizing this formula with respect to the separated or combined parameters is presented. The optimum values of the parameters are estimated iteratively at each pixel. The performance of the suggested approach is tested in the presence of noise with different levels and is compared to the traditional warped distance interpolation technique. A comparison of the suggested algorithm performance with other different interpolation techniques used in the commercial ACDSee Software is presented. The computational complexity of the suggested algorithm is also studied in the paper. The obtained results ensure the superiority of the suggested adaptive interpolation algorithm as compared to the traditional algorithms from both of the MSE and edge preservation points of view. As the results imply, the computation time of the suggested algorithm is moderate.
The same object seen in two different images can be geometrically and photometrically transformed. In this paper, a method of interest point detection and matching is described for the same object in different images. One of the main considerations is the change in the object scale. In this method, a reference scale is assigned to a particular instance of the object, and the change of scale is represented by a relative scale. Then, Harris' relative scale method is used for interest point detection. This method is robust to linear geometric transformations. A heuristic method for threshold selection is also described for robustness to intensity changes in a cluttered environment with partial occlusions. The repeatability rate of interest points for this method is higher then that for the existing methods. For the matching process, a local invariant descriptor is computed in the relative scale for each of the detected interest points. A hashing technique is applied to find the matches efficiently. The matching method enables finding a good number of correct matches for different types of transformations in a cluttered environment and one with partial occlusions. The proposed single scale detection and matching method could be effectively used for many practical applications, where the relative scale of the object can be predicted in advance.
Key words: relative scale, interest point, local invariant descriptor, repeatability rate, moment invariant, information content, multidimensional indexing.
Koprowski R., Wrobel Z.:
Automatic segmentation of biological cell structures based on conditional opening and closing.
MGV vol. 14, no. 3, 2005, pp. 285-307.
In this work we present an innovative algorithm for the automatic segmentation of biological cell structures based on two morphological operations - conditional opening or conditional closing. The operations do not have all the properties of classic erosion and dilation operations, in particular the properties of opening and closing. The metrological properties of the devised algorithm are presented, with an emphasis on measurement errors that result from the method used for the operation parameter selection.
Sensitivity of the algorithm to additive noise of Gaussian distribution is studied, taking into special consideration the microscopic images obtained from the examination of biological cell structures. We also show the convergence of the devised algorithm and its sensitivity to the quality of image binarization.
Some examples of segmentation of biological cell structure images obtained from implementations of the algorithm, both in the Matlab application with the Image Processing Toolbox and in Delphi, are shown.
Pattern recognition based on homology theory.
MGV vol. 14, no. 3, 2005, pp. 309-324.
This paper describes a method for comparing and recognizing patterns based on homology theory. We present algorithms and an exemplary application to handwritten letter recognition but the proposed idea can be easily used to recognize patterns in any dimension.
Key words: homology algorithm, size function, handwritten letter recognition.
Hussain M., Okada Y.:
LOD modelling of polygonal models.
MGV vol. 14, no. 3, 2005, pp. 325-343.
An automatic edge-collapse based simplification method has been proposed for decimation of polygonal models and generating their LODs (Levels of detail). The measure of geometric fidelity employed is motivated by the normal space deviation of a polygonal model arising during its decimation process and forces the algorithm to minimize the normal space deviation. In spite of the global nature of the evaluation of geometric deviation, the algorithm is memory efficient and involves less execution time then the state-of-the art simplification algorithms. This automatically prevents the creation of folds and automatically preserves visually important features of the model even at low levels of detail. LODs generated by our method compare favorably with those produced by the standard QEM-based algorithm QSlim in terms of the mean and maximum geometric errors, whereas its performance in preserving normal space of the original model is better than that of QSlim.
Key words: triangular meshes, surface simplification, level of detail, edge-collapse, shape approximation, multiresolution modeling.
New Books Notes
MGV vol. 14, no. 3, 2005, p. 345.
[A. Korzynska, M. Przytulska:
Przetwarzanie obrazów - Cwiczenia. (Image Processing - Exercises, in Polish).
Published by PJWSTK, Warsaw 2005.]
Hernández Hoyos M., Orkisz M., Douek P.C., Magnin I.E.:
Assessment of carotid artery stenoses in 3D contrast-enhanced magnetic resonance angiography, based on improved generation of the centerline.
MGV vol. 14, no. 4, 2005, pp. 349-378.
A method is proposed for generation of the centerline of 3D tubular shapes using an extensible-skeleton model. Starting from a user-selected point, the skeleton grown by iteratively adding subsequent centerline points within a prediction-estimation scheme controlled by a multi-scale analysis of the image moments. The location of the next point is predicted according the local orientation of the tubular structure. The coordinates of the predicted point are corrected under the influence of image forces and of prior model shape constraints. The extraction of artery centerlines from magnetic resonance angiography (MRA) images is described. The goal is a quantitative assessment of arterial stenoses based on cross-sectional diameters and areas of the vessel contours in the planes locally perpendicular to the centerline. For this purpose, iso-contours extraction based on an adaptive local iso-value have been implemented. The robustness and accuracy of the method have been demonstrated on MRA data on 5 reference phantoms and on 17 patients' carotid arteries. 97% of the centerlines were exploitable in the carotid arteries (100% in the phantoms). On average, the centerlines were extracted within 1 second, and the whole quantification process took less than 1 minute per artery, including interaction and display. The Mean difference (± standard deviation) between stenosis percentages, semi-automatically measured and visually estimated by radiologists, was 0.23% ± 7.89%. The reproducibility of the semi-automatic method was significantly better.
Key words: 3D centerline, active model, 3D moments, angiography.
Jankó Z., Kós G., Chetverikov D.:
Creating entirely textured 3D models of real objects using surface flattening.
MGV vol. 14, no. 4, 2005, pp. 379-398.
We present a novel method to create entirely textured 3D models of real objects by combiningpartial texture mappings using surface flattening (surface parametrisation). Texturing a 3D model is not trivial. Texture mappings can be obtained from optical images, but usually one imageis not sufficient to show the whole object; multiple images are required to cover the surface entirely. Merging partial texture mappings in 3D is difficult. Surface flattening converts a 3D mesh into 2D space preserving its structure. Transforming optical images to flattening-based texture maps allows them to be merged based onthe structure of the mesh. In this paper we describe a novel method for merging texture mappings using flattening and show its results on synthetic and real data.
Key words: 3D modelling, texturemapping, flattening, surface parametrisation.
Maskey M., Newman T.S.:
Cumulus cloud synthetic rendering techniques and their evaluations.
MGV vol. 14, no. 4, 2005, pp. 399-425.
Three new techniques for synthesizing realistic renderings of cumulus clouds are introducedand evaluated. The techniques utilize variations of the Perlin Noise and Koch fractals to achieve a reasonable cloud-like shape and texture. To evaluate the quality of renderings produced by the techniques, three classes of texture features are considered using cluster quality measures. Rendering quality is also evaluated versus real images using shape and texture features.
Key words: Perlin noise, Koch curve, volume rendering, fractals, texture, rendering quality, content-based image retrieval.
Rahman M., Kaneda K., Harada K.:
A new topology-based watermarking method for layered 3D triangular mesh models.
MGV vol. 14, no. 4, 2005, pp. 427-439.
A new topology-based watermarking method is proposed to embedi nformation in objects with layered 3D triangular meshes such as those reconstructed from CT or MRI data. The main idea of the methodis to compare the heights of the vertices of a triangle lying in the same layer. A watermark message is converted into a binary bit sequence, and then embedded into the model in such a way that the first vertex of a triangle in the upper level carries information 1, and the first vertex of a triangle in the lower level carries information 0. For experimental purposes, a watermark message is embedded in a mouse embryo model. It is robust against translation, rotation, re-sectioning, local deformation and scaling. It lefts some artifacts after re-arrangement of local or global numbering. It is useful for shape sensitive 3D geometric models.
Key words: watermarking, layered 3D triangular mesh model, topology based embedding, computer graphics.
Yang L., Sahli H., Hào D.N.:
A variational approach to 3D line orientation estimation from motion.
MGV vol. 14, no. 4, 2005, pp. 441-453.
A variational approach to estimating 3D line orientation from motion is presented. A 2D motion constraint on 3D lines regularized by a quadratic term is used to set up an objective functional. From its associated Euler-Lagrange equations, we develop a vector-valued diffusion model, with a reaction term based on the 2D motion constraint. Three separate diffusion processes, corresponding to each component of the 3D line orientation, are coupled with each other through the reaction term and evolve simultaneously. Each 3D line orientation is estimated separately. The regularization parameter is estimated by an L-curve, which provides a better estimation. Experimental results from image sequences indicate stability and accuracy of the approach.
Key words: line orientation, motion, variational approach, vector-valued reaction diffusion, L-curve.
Geometrical Wavelets and their Generalizations in Digital Image Coding and Processing.
MGV vol. 14, no. 4, 2005, p. 455.
Contents of volume 14, 2005