Acoustic imaging is an important and active area of research in health-related fields, especially those related to human speech disorders. For speech and other acoustic signals, the sound spectrogram is an important application tool for examining sound intensity and spectral relationships over time. Spectrogram features, like those of other images, can often be described in terms of boundary and texture parameters, and this proposed project will explore the hypothesis that certain image processing techniques may be useful for describing and quantifying sound spectrogram contour information. Fourier Descriptors (FD's) are a special Fourier technique that have proved successful in analyzing 2- and 3-dimensional image boundaries, and the application of FD's to spectrogram components will be the principal technique examined. Previous research has indicated that a relatively small number of FD's are required to characterize a boundary, and clustering techniques are useful for maximizing confidence in the optimal number of Descriptors. In this study, cluster analysis will be used to examine results as test samples of known origin are classified as to group using nearest neighbor methods with the Euclidean distance metric and a varying number of FD's. Pattern recognition structural methods involving formal parsing algorithms and grammar-based replacement rules will be examined for capturing relationships among sound components. Test results will be obtained for five groups of sounds, including various bioacoustic sounds and human speech. By using the inverse FFT on FD's, this project should suggest methods for simulating sounds from a relatively small set of Fourier Descriptors. It is anticipated that the developed methodology can be integrated with human expertise into a larger, long-term project involving high-performance expert systems that combine symbolic processing with numerical methods for acoustic imaging.
