Aumento de resolução de imagens de ressonância magnética do trato vocal utilizadas em modelos de síntese articulatória

Abstract

Articulatory Synthesis consists in reproducing speech by means of models of the vocal tract and of articulatory processes. Recent advances in Magnetic Resonance Imaging (MRI) allowed for important improvements with respect to the speech comprehension and the forms taken by the vocal tract. However, one of the main challenges in the field is the fast and at the same time high-quality acquisition of image sequences. Since adopting more powerful acquisition devices might be financially inviable, a more feasible solution proposed in the literature is the resolution enhancement of the images by changes introduced in the acquisition model. This dissertation proposes a method for the spatio-temporal resolution enhancement of the obtained sequences using only digital image processing techniques. The approach involves two stages: (1) the temporal resolution enhancement by means of a motion compensated interpolation technique; and (2) the spatial resolution enhancement by means of a super resolution image reconstruction technique. With respect to the temporal resolution enhancement, two interpolation models are compared: linear interpolation considering two adjacent images and cubic splines interpolation considering four contiguous images. Since both models performed equally in the experiments, the linear interpolation was adopted, for its simplicity and lower computational cost. The initial goal of the spatial resolution enhancement was an extension of the candidate s approach proposed in her master s thesis. Adopting a maximum a posteriori probability approach (MAP), the high-resolution images were modeled using the Markov Random Fields (MRF) Generalized Isotropic Multi-Level Logistic (GIMLL) model and the Iterated Conditional Modes (ICM) algorithm. However, even though the approach has presented promising results, due to the dimension of the target problem, the algorithm presented high computational cost. Considering this limitation, an adaptation of the Wiener filter for the super-resolution reconstruction problem was considered. Inspired by two methods available in the literature, three approaches were proposed: the statistical interpolation, the multi-temporal approach, and the adaptive Wiener filter. In all cases, a separable Markovian model and an isotropic model were compared in the characterization of the spatial correlation structures. These models were used to characterize the correlation and cross correlation of observations for the statistical interpolation and the multi-temporal approach. On the other hand, for the adaptive Wiener filter, these models were used to characterize the a priori spatial correlation. According to the conducted experiments, the isotropic model outperformed the separable Markovian model. Besides, considering all Wiener filter-based approaches and the initial approach based on the GIMLL model, the adaptive Wiener filter outperformed all other approaches and was also faster than a single iteration of the GIMLL-based approach.