Catálogo de publicaciones - libros

In this paper we compare the overall image quality of 7 state-of-the-art denoising schemes, based on human visual perception. A psycho-visual experiment was set up in which 37 subjects were asked to score and compare denoised images. A perceptual space is constructed from this experiment through multidimensional scaling (MDS) techniques using the perceived dissimilarity and quality preference between the images and the scaled perceptual attributes bluriness and artefacts.

We found that a two-dimensional perceptual space adequately represents the processed images used in the experiment, and that the perceptual spaces obtained for all scenes are very similar. The interpretation of this space leads to a ranking of the filters in perceived overall image quality. We can show that the impairment vector, whose direction is opposite to that of the quality vector, lies between the attribute vectors for bluriness and artefacts, which on their account form an angle of about 35 degrees meaning they do interact. A follow-up experiment allowed us to determine even further why subjects preferred one filter over the other.

A new multidimensional modeling of data has recently been introduced, which can be used a wide range of signals. This paper presents multiway filtering for denoising hyperspectral images. This approach is based on a tensorial modeling of the desired information. The optimization criterion used in this multiway filtering is the minimization of the mean square error between the estimated signal and the desired signal. This minimization leads to some estimated -mode filters which can be considered as the extension of the well-known Wiener filter in a particular mode. An ALS algorithm is proposed to determine each -mode Wiener filter. Using the ALS loop allows to take into account the mode interdependence. This algorithm requires the signal subspace estimation for each mode. In this study, we have extended the well-know Akaike Information Criterion (AIC) and the minimum description length (MDL) criterion to detect the number of dominant eigenvalues associated with the signal subspace. The performance of this new method is tested on hyperspectral images. Comparative studies with classical bidimensional filtering methods show that our algorithm presents good performances.

Image segmentation using graph cuts have become very popular in the last years. These methods are computationally expensive, even with hard constraints (seed pixels). We present a solution that runs in time proportional to the number of pixels. Our method computes an ordered region growing from a set of seeds inside the object, where the of each pixel is proportional to the cost of an in the image graph from the seed set to that pixel. Each pixel defines a region which includes it and all pixels with lower propagation order. The boundary of each region is a possible cut boundary, whose cut measure is also computed and assigned to the corresponding pixel on-the-fly. The object is obtained by selecting the pixel with minimum-cut measure and all pixels within its respective cut boundary. Approaches for graph-cut segmentation usually assume that the desired cut is a global minimum. We show that this can be only verified within a reduced search space under certain hard constraints. We present and evaluate our method with three cut measures: normalized cut, mean cut and an energy function.

A new design for image processing frameworks is proposed. The new design addresses high-level abstractions suited for component-based image processing applications, in particular real-time image processing with high performance demands. The RIM framework, an implementation of this design, is gone through. It is explained how RIM can be adapted in applications, and integrated with other image libraries. It is also shown how it can be used to confirm some properties of widely used image formats.

Only three set of pattern chain elements to detect corners in irregular shapes are introduced. A code based on three orthogonal change directions, when visiting a contour shape, are used. Previous approaches for detecting corners employ eight different symbols and usually compute angles and maximum curvature. The three basic pattern contour chain elements, founded in this paper, represent changes of direction in the contour curves, requiring few computing power to obtain corners. Also, we have found that the method is independent of shape orientation.

We propose a novel active contour model by incorporating particle based electrostatic interactions into the geometric active contour framework. The proposed active contour, embedded in level sets, propagates under the joint influence of a boundary attraction force and a boundary competition force. Unlike other contour models, the proposed vector field dynamically adapts by updating itself when a contour reaches a boundary. The model is then more invariant to initialisation and possesses better convergence abilities. Analytical and comparative results are presented on synthetic and real images.

Recent developments in dynamic contour tracking in video sequences are based on prediction using dynamical models. The parameters of these models are fixed by learning the dynamics from a training set to represent plausible motions, such as constant velocity or critically damped oscillations. Thus, a problem arise in cases of non-constant velocity and unknown interframe motion, unlearned motions, and the CONDENSATION algorithm fails to track the dynamic contour. The main contribution of this work is to propose an adaptative dynamical model which parameters are based on non-linear/non-gaussian observation models. We study two different approaches, one statistical and one shape-based, to estimate the deformation of an object and track complex dynamics without learning from a training set neather the dynamical nor the deformation models and under the constraints of missing data, non-linear deformation and unknown interframe motion. The developed approaches have been successfully tested on several sequences.

The use of active contours for texture segmentation seems rather attractive in the recent research, indicating that such methodologies may provide more accurate results. In this paper, a novel model for texture segmentation is presented, combining advantages of the active contour approach with texture information acquired by the Local Binary Pattern (LBP) distribution. The proposed LBP scheme has been formulated in order to capture regional information extracted from distributions of LBP values, characterizing a neighborhood around each pixel, instead of using a single LBP value to characterize each pixel. The log-likelihood statistic is employed as a similarity measure between the LBP distributions, resulting to more detailed and accurate segmentation of texture images.

An approach is presented for characterize objects and image texture by local lacunarity. This measure makes possible to distinguish sets that have same fractal dimension. In image analysis it can be used as a new feature in the pattern recognition process mainly for identification of natural textures. Illustrating the approach, two types of examples were presented: 3D objects representing approximations of fractal sets and medical images. In the first type, we apply this approach to show its possibility when the objects presents the same fractal dimension. The second type shows that it can be used as a feature on pattern recognition alone in many resolutions.

In general, the less probable an event, the more attention we pay to it. Likewise, considering visual perception, it is interesting to regard important image features as those that most depart from randomness. This statistical approach has recently led to the development of adaptive and parameterless algorithms for image analysis. However, they require computer-intensive statistical measurements. Digital retinas, with their massively parallel and collective computing capababilities, seem adapted to such computational tasks. These principles and opportunities are investigated here through a case study: extracting meaningful segments from an image.