Recent advances in real-time magnetic resonance imaging (rtMRI) of the vocal tract provides opportunities for studying human speech. This modality together with acquired speech may enable the mapping of articulatory configurations to acoustic features. In this study, we take the first step by training a deep learning model to classify 27 different phonemes from midsagittal MR images of the vocal tract.An American English database was used to train a convolutional neural network for classifying vowels (13 classes), consonants (14 classes) and all phonemes (27 classes) of 17 subjects. Classification top-1 accuracy of the test set for all phonemes was 57%. Erroranalysis showedvoiced and unvoiced sounds often being confused. Moreover, we performed principal component analysis on the network’s embedding and observed topological similarities between thenetwork learned representation and the vowel diagram.Saliency maps gaveinsight intothe anatomical regions most important for classification and show congruence with knownregions of articulatory importance.We demonstrate the feasibility for deep learning to distinguish between phonemes from MRI. Network analysis can be used to improve understanding of normal articulation and speech and, in the future, impaired speech. This study brings us a step closer to the articulatory-to-acoustic mapping from rtMRI.

An "inconsistent" particle filter produces--in a statistical sense--larger estimation errors than predicted by the model on which the filter is based. Two test variables are introduced that allow the detection of inconsistent behavior. The statistical properties of the variables are analyzed. Experiments confirm their suitability for inconsistency detection.

There are two open problems when finite mixture densities are used to model multivariate data: the selection of the number of components and the initialization. In this paper, we propose an online (recursive) algorithm that estimates the parameters of the mixture and that simultaneously selects the number of components. The new algorithm starts with a large number of randomly initialized components. A prior is used as a bias for maximally structured models. A stochastic approximation recursive learning algorithm is proposed to search for the maximum a posteriori (MAP) solution and to discard the irrelevant components.

We propose a technique that combines geometric hashing with stereo vision. The idea is to use the robustness of geometric hashing to spurious data to overcome the correspondence problem, while the stereo vision setup enables direct model matching using the 3-D object models. Furthermore, because the matching technique relies on the relative positions of local features, we should be able to perform robust recognition even with partially occluded objects. We tested this approach with simple geometric objects using a corner point detector. We successfully recognized objects even in scenes where the objects were partially occluded by other objects. For complicated scenes, however, the limited set of model features and required amount of computing time, sometimes became a problem.

Image segmentation based on contour extraction usually involves three stages of image operations: feature extraction, edge detection and edge linking. This paper is devoted to the first stage: a method to design feature extractors used to detect edges from noisy and/or blurred images. The method relies on a model that describes the existence of image discontinuities (e.g. edges) in terms of covariance functions. The feature extractor transforms the input image into a "log-likelihood ratio" image. Such an image is a good starting point of the edge detection stage since it represents a balanced trade-off between signal-to-noise ratio and the ability to resolve detailed structures. For 1-D signals, the performance of the edge detector based on this feature extractor is quantitatively assessed by the so called "average risk measure". The results are compared with the performances of 1-D edge detectors known from literature. Generalizations to 2-D operators are given. Applications on real world images are presented showing the capability of the covariance model to build edge and line feature extractors. Finally it is shown that the covariance model can be coupled to a MRF-model of edge configurations so as to arrive at a maximum a posteriori estimate of the edges or lines in the image.>

A new method for evaluation of edge detectors, based on the average risk of a decision, is discussed. The average risk is a performance measure well-known in Bayesian decision theory. Since edge detection can be regarded as a compound decision making process, the performance of an edge detector is context dependent. Therefore, the application of average risk to edge detection is non-trivial. The paper describes a method to estimate the probabilities on a number of different types of (context dependent) errors. A weighted sum of these estimated probabilities represents the average risk. The weight coefficients define the cost function. The method is suitable, not only for the comparison of edge operators, but also for the determination of the weaknesses and strengths of a certain edge operator. This is demonstrated with some well-known edge operators.>