Carmen Hernández

Two lattice computing approaches for the unsupervised segmentation of hyperspectral images

Endmembers for the spectral unmixing analysis of hyperspectral images are sets of affinely independent vectors, which define a convex polytope covering the data points that represent the pixel image spectra. Strong lattice independence (SLI) is a property defined in the context of lattice associative memories convergence analysis. Recent results show that SLI implies affine independence, confirming the value of lattice associative memories for the study of endmember induction algorithms. In fact, SLI vector sets can be easily deduced from the vectors composing the lattice auto-associative memories (LAM). However, the number of candidate endmembers found by this algorithm is very large, so that some selection algorithm is needed to obtain the full benefits of the approach. In this paper we explore the unsupervised segmentation of hyperspectral images based on the abundance images computed, first, by an endmember selection algorithm and, second, by a previously proposed heuristically defined algorithm. We find their results comparable on a qualitative basis.

Lattice independent component analysis for functional magnetic resonance imaging

We introduce a lattice independent component analysis (LICA) unsupervised scheme to functional magnetic resonance imaging (fMRI) data analysis. LICA is a non-linear alternative to independent component analysis (ICA), such that ICAs statistical independent sources correspond to LICAs lattice independent sources. In this paper, LICA uses an incremental lattice source induction algorithm (ILSIA) to induce the lattice independent sources from the input dataset. The ILSIA computes a set of Strongly Lattice Independent vectors using properties of lattice associative memories regarding Lattice Independence and Chebyshev best approximation. The lattice independent sources constitute a set of Affine Independent vectors that define a simplex covering the input data. LICA carries out data linear unmixing based on the lattice independent sources basis. Therefore, LICA is a hybrid combination of a non-linear lattice based component and a linear unmixing component. The principal advantage over ICA is that LICA does not impose any probabilistic model assumptions on the data sources. We compare LICA with ICA in two case studies. Firstly, including simulated fMRI data, LICA discovers the spatial location of meaningful sources with less ambiguity than ICA. Secondly, including real data from an auditory stimulation experiment, LICA improves over some state of the art ICA variants discovering the activation patterns detected by Statistical Parametric Mapping (SPM) on the same data.

Enhancing the Feature Extraction Process for Automatic Speech Recognition with Fractal Dimensions

Mel frequency cepstral coefficients (MFCCs) are a standard tool for automatic speech recognition (ASR), but they fail to capture part of the dynamics of speech. The nonlinear nature of speech suggests that extra information provided by some nonlinear features could be especially useful when training data are scarce or when the ASR task is very complex. In this paper, the Fractal Dimension of the observed time series is combined with the traditional MFCCs in the feature vector in order to enhance the performance of two different ASR systems. The first is a simple system of digit recognition in Chinese, with very few training examples, and the second is a large vocabulary ASR system for Broadcast News in Spanish.

A single individual evolutionary strategy for endmember search in hyperspectral images

We define an single individual evolutionary strategy (SIE) for the induction of a set of endmembers and we compare it with a conventional evolutionary strategy (ES), tailored to this task, over a couple of hyperspectral images. The SIE considers a set of end-members as an evolving population. Individuals correspond to hypothetical endmember spectra, and they are selected as candidates for mutation on the basis of their partial abundance images. The populations global fitness when the mutated individual substitutes its parent is the measure of the goodness of tile individual. Although the aim of defining the SIE was to reduce the computational cost of applying ES to the high volume data in hyperspectral images, we have found that SIE also improves the fitness performance of conventional ES.

Further results on swarms solving graph coloring

We have proposed the mapping of graph coloring problems into swarm dynamics. Empirical evidence that flock steering behaviors augmented with the notion of hostility (enmity and friendliness) are enough to perform efficiently the task of coloring the nodes of graphs even in the case 3-coloration hard graph topologies. We discuss here what are the minimal cognitive capabilities that allow the emergent behavior of swarms to solve such NP-complete problem without mediating an explicit knowledge representation.

A Morphological Cellular Automata Based On Morphological Independence

We discuss a definition of Morphological Cellular Neural Networks (MCNN) where the state change operator are Auto-associative Morphological Memories (AMM). The fast convergence properties of AMM and the shape of its fixed point set make the MCNN dynamics trivial. However, segmentation results are poor. We propose a Morphological Cellular Automata (MCA) with assured convergence to a state characterized by morphological dependences and independences between neighbouring cell states. Cell dynamic rules test morphological dependence among neighbouring cell’s states. When neighbouring cell states are morphological dependent in the erosive or dilative sense, the morphologically dominant state colonizes the neighbour with morphological dependent state. The resulting configuration of cell states is composed of homogeneous regions whose boundaries are defined by the morphological independence relation. Results are given on image segmentation, where MCA cells correspond to image pixels.

Further results on AMM for endmember induction

Our main interest is to perform unsupervised segmentation of the hyperspectral images. Our approach is to interpret abundance images resulting from spectral unmixing as the characterization of regions in the image. We induce the endmembers needed for spectral unmixing from the image data. Therefore the endmember spectra are not easily interpretable as laboratory spectra. Our method for endmember induction looks at the morphological independence or the endmembers as a necessary condition. We use the Autoassociative Morphological Memories (AMM) as detectors of morphological independence conditions. Our algorithm needs only one pass of the image. The experimental results obtained over a set of synthetic images are presented here, contrasted with the ICA and CCA approaches.

Data fusion for high spatial resolution LAI estimation

Leaf Area Index (LAI) is a critical variable for forest management. It is difficult to obtain accurate LAI estimations of high spatial resolution over large areas. Local estimations can be obtained from in situ field measurements. Extrapolation of local measurements is prone to error. Remote sensing LAI estimation products, such as the one provided by MODIS are of very low resolution and subject to criticism in recent validation works. Forest management requires increasingly high resolution estimations of LAI. We propose a data fusion process for high spatial resolution estimation of the LAI over a large area, combining several heterogeneous information sources: field sampled data, elevation data and remote sensing data. The process makes use of spatial interpolation techniques. We follow a hybrid validation approach that combines the conventional prediction error measures with a spatial validation based on image segmentation. We obtain encouraging results of this information fusion process on data from a forest area in the north of Portugal.

Combining Mel frequency Cepstral coefficients and fractal dimensions for automatic speech recognition

Hidden Markov Models and Mel Frequency Cepstral Coefficients (MFCCs) are a sort of standard for Automatic Speech Recognition (ASR) systems, but they fail to capture the nonlinear dynamics of speech that are present in the speech waveforms. The extra information provided by the nonlinear features could be especially useful when training data is scarce, or when the ASR task is very complex. In this work, the Fractal Dimension (FD) of the observed time series is combined with the traditional MFCCs in the feature vector in order to enhance the performance of two different ASR systems: the first one is a very simple one, with very few training examples, and the second one is a Large Vocabulary Continuous Speech Recognition System for Broadcast News.

On the use of a hybrid approach to contrast endmember induction algorithms

In remote sensing hyperspectral image processing, identifying the constituent spectra (endmembers) of the materials in the image is a key procedure for further analysis The contrast between Endmember Inductions Algorithms (EIAs) is a delicate issue, because there is a shortage of validation images with accurate ground truth information, and the induced endmembers may not correspond to any know material, because of illumination and atmospheric effects In this paper we propose a hybrid validation method, composed on a simulation module which generates the validation images from stochastic models and evaluates the EIA through Content Based Image Retrieval (CBIR) on the database of simulated hyperspectral images We demonstrate the approach with two EIA selected from the literature.