Luis Pizarro

Generalised Nonlocal Image Smoothing

We propose a discrete variational approach for image smoothing consisting of nonlocal data and smoothness constraints that penalise general dissimilarity measures defined on image patches. One of such dissimilarity measures is the weighted L2 distance between patches. In such a case we derive an iterative neighbourhood filter that induces a new similarity measure in the photometric domain. It can be regarded as an extended patch similarity measure that evaluates not only the patch similarity of two chosen pixels, but also the similarity of their corresponding neighbours. This leads to a more robust smoothing process since the pixels selected for averaging are more coherent with the local image structure. By slightly modifying the way the similarities are computed we obtain two related filters: The NL-means filter of Buades etxa0al. (SIAM Multiscale Model. Simul. 4(2):490–530, 2005b) and the NDS filter of Mrázek etxa0al. (Geometric Properties for Incomplete Data, Computational Imaging and Vision, vol.xa031, pp.xa0335–352, Springer, Dordrecht, 2006). In fact, the proposed approach can be considered as a generalisation of the latter filter to the space of patches. We also provide novel insights into relations of the NDS filter with diffusion/regularisation methods as well as with some recently proposed graph regularisation techniques. We evaluate our method for the task of denoising greyscale and colour images degraded with Gaussian and salt-and-pepper noise, demonstrating that it compares very well to other more sophisticated approaches.

Robust automated multiple view inspection

Recently, Automated Multiple View Inspection (AMVI) has been developed for automated defect detection of manufactured objects, and the framework was successfully implemented for calibrated image sequences. However, it is not easy to be implemented in industrial environments because the calibration is a difficult and an unstable process. To overcome these disadvantages, the robust AMVI strategy, which assumes that an unknown affine transformation exists between each pair of uncalibrated images, is proposed. This transformation is estimated using two complementary robust procedures: a global approximation of the affine mapping is computed by creating candidate correspondences via B-splines and selecting those which better satisfy the epipolar constraint for uncalibrated images. Then, we use this approximation as initial estimate of a robust intensity-based matching approach, which is applied locally on each potential defect. The result is that false alarms are discarded, and the defects of an industrial object are actually tracked along the uncalibrated image sequence. The method is successful as shown in our experiments on aluminum die castings.

Adaptive Continuous-Scale Morphology for Matrix Fields

In this article we consider adaptive, PDE-driven morphological operations for 3D matrix fields arising e.g. in diffusion tensor magnetic resonance imaging (DT-MRI). The anisotropic evolution is steered by a matrix constructed from a structure tensor for matrix valued data. An important novelty is an intrinsically one-dimensional directional variant of the matrix-valued upwind schemes such as the Rouy-Tourin scheme. It enables our method to complete or enhance anisotropic structures effectively. Axa0special advantage of our approach is that upwind schemes are utilised only in their basic one-dimensional version, hence avoiding grid effects and leading to an accurate algorithm. No higher dimensional variants of the schemes themselves are required. Experiments with synthetic and real-world data substantiate the gap-closing and line-completing properties of the proposed method.

A Directional Rouy-Tourin Scheme for Adaptive Matrix-Valued Morphology

In order to describe anisotropy in image processing models or physical measurements, matrix fields are a suitable choice. In diffusion tensor magnetic resonance imaging (DT-MRI), for example, information about the diffusive properties of water molecules is captured in symmetric positive definite matrices. The corresponding matrix field reflects the structure of the tissue under examination. Recently, morphological partial differential equations (PDEs) for dilation and erosion known for grey scale images have been extended to matrix-valued data. n nIn this article we consider an adaptive, PDE-driven dilation process for matrix fields. The anisotropic morphological evolution is steered with a matrix constructed from a structure tensor for matrix valued data. An important novel ingredient is a directional variant of the matrix-valued Rouy-Tourin scheme that enables our method to complete or enhance anisotropic structures effectively. Experiments with synthetic and real-world data substantiate the gap-closing and line-completing properties of the proposed method.

Bimodal biometric person identification system under perturbations

Multibiometric person identification systems play a crucial role in environments where security must be ensured. However, building such systems must jointly encompass a good compromise between computational costs and overall performance. These systems must also be robust against inherent or potential noise on the data-acquisition machinery. In this respect, we proposed a bimodal identification system that combines two inexpensive and widely accepted biometric traits, namely face and voice information. We use a probabilistic fusion scheme at the matching score level, which linearly weights the classification probabilities of each person-class from both face and voice classifiers. The system is tested under two scenarios: a database composed of perturbation-free faces and voices (ideal case), and a database perturbed with variable Gaussian noise, salt-and-pepper noise and occlusions. Moreover, we develop a simple rule to automatically determine the weight parameter between the classifiers via the empirical evidence obtained from the learning stage and the noise level. The fused recognition systems exceeds in all cases the performance of the face and voice classifiers alone.

PDE-Driven Adaptive Morphology for Matrix Fields

Matrix fields are important in many applications since they are the adequate means to describe anisotropic behaviour in image processing models and physical measurements. A prominent example is diffusion tensor magnetic resonance imaging (DT-MRI) which is a medical imaging technique useful for analysing the fibre structure in the brain. Recently, morphological partial differential equations (PDEs) for dilation and erosion known for grey scale images have been extended to three dimensional fields of symmetric positive definite matrices. n nIn this article we propose a novel method to incorporate adaptivity into the matrix-valued, PDE-driven dilation process. The approach uses a structure tensor concept for matrix data to steer anisotropic morphological evolution in a way that enhances and completes line-like structures in matrix fields. Numerical experiments performed on synthetic and real-world data confirm the gap-closing and line-completing qualities of the proposed method.

Evaluating a general class of filters for image denoising

Recently, an energy-based unified framework for image denoising was proposed by Mrazek et al. [10], from which existing nonlinear filters such as M-smoothers, bilateral filtering, diffusion filtering and regularisation approaches, are obtained as special cases. Such a model offers several degrees of freedom (DOF) for tuning a desired filter. In this paper, we explore the generality of this filtering framework in combining nonlocal tonal and spatial kernels. We show that Bayesian analysis provides suitable foundations for restricting the parametric space in a noisedependent way. We also point out the relations among the distinct DOF in order to guide the selection of a combined model, which itself leads to hybrid filters with better performance than the previously mentioned special cases. Moreover, we show that the existing trade-off between the parameters controlling similarity and smoothness leads to similar results under different settings.

A Generic Neighbourhood Filtering Framework for Matrix Fields

The Nonlocal Data and Smoothness (NDS) filtering framework for greyvalue images has been recently proposed by Mrazek et al. This model for image denoising unifies M-smoothing and bilateral filtering, and several well-known nonlinear filters from the literature become particular cases. In this article we extend this model to so-called matrix fields. These data appear, for example, in diffusion tensor magnetic resonance imaging (DT-MRI). Our matrix-valued NDS framework includes earlier filters developped for DT-MRI data, for instance, the affine-invariant and the log-Euclidean regularisation of matrix fields. Experiments performed with synthetic matrix fields and real DT-MRI data showed excellent performance with respect to restoration quality as well as speed of convergence.

Robust Estimation of Roughness Parameter in SAR Amplitude Images

The precise knowledge of the statistical properties of synthetic aperture radar (SAR) data plays a central role in image processing and understanding. These properties can be used for discriminating types of land uses and to develop specialized filters for speckle noise reduction, among other applications. In this work we assume the distribution (mathcal{G}^{0}_{A}) as the universal model for multilook amplitude SAR images under the multiplicative model. We study some important properties of this distribution and some classical estimators for its parameters, such as Maximum Likelihood (ML) estimators, but they can be highly influenced by small percentages of ‘outliers’, i.e., observations that do not fully obey the basic assumptions. Hence, it is important to find Robust Estimators. One of the best known classes of robust techniques is that of M estimators, which are an extension of the ML estimation method. We compare those estimation procedures by means of a Monte Carlo experiment.