|
Descripción
|
|
|---|---|
| Joint representations have experimented a significant height in signal processing during the last decades, to such an extent that there is no topic they have not been utilized for. Within a sea of joint representations existent in the literature, one of them concerns the present work: the log-Gabor multiresolution transform proposed in [70, 68]. It has low spectral overlapping, the scalability and high selectivity in orientation, shiftinvariance, self-invertibility and complex definition, what confers efficiency, versatility and robustness against noise and a low presence of artifacts. Further on, the tight similarity of overcomplete log-Gabor filters to the cortical area V1, together with the modeling of inhibitory/facilitatory neuronal behaviors and sparse coding algorithms allow to achieve an approximation of the image based on the extraction of those salient features normally coincident with contours. This type of image representation based on multiscale contours traces new routes to solve image processing tasks, in particular the areas of image compression and fusion. A recent compression paradigm postulates higher efficiency from coding separately features present in images, such as luminance, contours or textures. Following that paradigm, in this thesis a new compression method is proposed based on coding those multiscale features extracted from the sparse log-Gabor transformation. In account of the nature of such features, a chain coding algorithm has been specially tailored to the stochastic and morphological peculiarities of multiscale contours. Thus, different predictive techniques as well as prefix and arithmetic coding has been combined according to each alphabet. Moreover, the proposed algorithm offers a complete compression scheme including low-pass coding as well as header bitstream allocation. Such coding rest on a model of the primary visual cortex in order to mitigate typical compression distortions usually produced by compression standards such as JPEG and JPEG2000. Multiresolution decompositions have proven their superiority against other traditional image fusion techniques, nevertheless it does not exist any evident hegemony, often due to the lack of a reference image. In this thesis, several types of wavelets were compared to log-Gabor filters, never used before due to its traditional lack of exact reconstruction, which succeeded remarkably. Further, a general algorithm for multiresolution schemes named multisize windows is proposed, which adapts the size of the averaging window according to the local features in the image, in contrast to traditional fixed window approaches. It exploits the advantages of both small, i.e. precise, and big, i.e. robust, windows showing significant reduction on errors in decision maps. Finally, a novel contour-based fusion method is proposed by integrating the multiscale contours to multiresolution fusion. This feature-based algorithm reduces the sensitivity to noise, blurring effects and misregistration artifacts. | |
|
Internacional
|
Si |
|
ISBN
|
|
|
Tipo de Tesis
|
|
|
Calificación
|
Sobresaliente cum laude |
|
Fecha
|
20/04/2007 |