Catálogo de publicaciones - libros

Deformational plagiocephaly is a term describing cranial a-symmetry and deformation commonly seen in infants. The purpose of this work was to develop a methodology for assessment and modelling of head asymmetry. The clinical population consisted of 38 infants for whom 3-dimensional surface scans of the head had been obtained both before and after their helmet orthotic treatment. Non-rigid registration of a symmetric template to each of the scans provided detailed point correspondence between scans. A new asymmetry measure was defined and was used in order to quantify and localize the asymmetry of each infant’s head, and again employed to estimate the improvement of asymmetry after the helmet therapy. A statistical model of head asymmetry was developed (PCA). The main modes of variation were in good agreement with clinical observations, and the model provided an excellent and instructive quantitative description of the asymmetry present in the dataset.

In this paper we propose a new approach to real-time view-based object recognition and scene registration. Object recognition is an important sub-task in many applications, as e.g., robotics, retrieval, and surveillance. Scene registration is particularly useful for identifying camera views in databases or video sequences. All of these applications require a fast recognition process and the possibility to extend the database with new material, i.e., to update the recognition system online.

The method that we propose is based on P-channels, a special kind of information representation which combines advantages of histograms and local linear models. Our approach is motivated by its similarity to information representation in biological systems but its main advantage is its robustness against common distortions as clutter and occlusion. The recognition algorithm extracts a number of basic, intensity invariant image features, encodes them into P-channels, and compares the query P-channels to a set of prototype P-channels in a database. The algorithm is applied in a cross-validation experiment on the COIL database, resulting in nearly ideal ROC curves. Furthermore, results from scene registration with a fish-eye camera are presented.

This paper introduces a new concept within shadow segmentation for usage in shadow removal and augmentation through construction of a multiplicity alpha overlay shadow model. Previously, an image was considered to consist of shadow and non-shadow regions. This makes it difficult to seamlessly remove shadows and insert augmented shadows that overlap real shadows. We construct a model that accounts for sunlit, umbra and penumbra regions by estimating the degree of shadow. The model is based on theories about color constancy, daylight, and the geometry that causes penumbra. A graph cut energy minimization is applied to estimate the alpha parameter. Overlapping shadow augmentation and removal is also demonstrated. The approach is demonstrated on natural complex image situations. The results are convincing, and the quality of augmented shadows overlapping real shadows and removed shadows depends on the quality of the estimated alpha gradient in penumbra.

Direct image registration methods usually treat shadows as outliers. We propose a method which registers images in a 1D shadow invariant space. Shadow invariant image formation is possible by projecting color images, expressed in a log-chromaticity space, onto an ‘intrinsic line’. The slope of the line is a camera dependent parameter, usually obtained in a prior calibration step. In this paper, calibration is avoided by jointly determining the ‘invariant slope’ with the registration parameters. The method deals with images taken by different cameras by using a different slope for each image and compensating for photometric variations. Prior information about the camera is, thus, not required. The method is assessed on synthetic and real data.

We use the popular active appearance models (AAM) for extracting discriminative features from images of biological objects. The relevant discriminative features are combined principal component (PCA) vectors from the AAM and texture features from cooccurrence matrices. Texture features are extracted by extending the AAM’s with a textural warp guided by the AAM shape. Based on this, texture cooccurrence features are calculated. We use the different features for classifying the biological objects to species using standard classifiers, and we show that even though the objects are highly variant, the AAM’s are well suited for extracting relevant features, thus obtaining good classification results. Classification is conducted on two real data sets, one containing various vegetables and one containing different species of wood logs.

In medical imaging, signals acquired in non-Cartesian coordinate systems are common. For instance, CT and MRI often produce significantly higher resolution within scan planes, compared to the distance between two adjacent planes. Even oblique sampling occurs, by the use of gantry tilt. In ultrasound imaging, samples are acquired in a polar coordinate system, which implies a spatially varying metric.

In order to produce a geometrically correct image, signals are generally resampled to a Cartesian coordinate system. This paper concerns estimation of local structure tensors directly from the non-Cartesian coordinate system, thus avoiding deteriorated signal and noise characteristics caused by resampling. In many cases processing directly in the warped coordinate system is also less time-consuming.

A geometrically correct tensor must obey certain transformation rules originating from fundamental differential geometry. Subsequently, this fact also affects the tensor estimation. As the local structure tensor is estimated using filters, a change of coordinate system also change the shape of the spatial support of these filters. Implications and limitations brought on by sampling require the filter design criteria to be adapted to the coordinate system.

The purpose of this research was to examine the learning system for a feature extraction unit in OCR. Average Risk Criterion and Probabilistic Descent (basic model of MCE/GPD) are employed in the character recognition system which consists of feature extraction with filters and Euclidian distance. The learning process was applied to the similar character discrimination problem and the effects were shown as the accuracy improvement.

In this paper a robust method is presented to classify and estimate an objects pose from a real time range image and a low dimensional model. The model is made from a range image training set which is reduced dimensionally by a nonlinear manifold learning method named Local Linear Embedding (LLE). New range images are then projected to this model giving the low dimensional coordinates of the object pose in an efficient manner. The range images are acquired by a state of the art SwissRanger SR-3000 camera making the projection process work in real-time.

The availability of ground-truth flow field is crucial for quantitative evaluation of any optical flow computation method. The fidelity of test data is also important when artificially generated. Therefore, we generated an artificial flow field together with an artificial image sequence based on real-world sample image. The presented framework benefits of a two-layered approach in which user-selected foreground was locally moved and inserted into an artificially generated background. The background is visually similar to input sample image while the foreground is extracted from original and so is the same. The framework is capable of generating 2D and 3D image sequences of arbitrary length. Several examples of the version tuned to simulate real fluorescent microscope images are presented. We also provide a brief discussion.