| The general problem that we consider is the real-time enhancement of a noisy image as it arrives column by column (time is associated with the horizontal axis). The enhancement is modeled by a column of real values between 0 and 1 (with as many pixels as in the noisy image column), which is used as a mask on the noisy image column. That is, the clean image column is estimated by multiplying the noisy image column by the mask column, pixel by pixel. In this work, the real-valued column is attained by "fuzzifying" a binary column, which is estimated from the noisy image column using a prior model and a noise model. The prior model is specified by a statistical distribution that assigns to binary columns a probability of occurrence. The noise model is specified by a statistical distribution that assigns the likelihood of occurrence of the noisy image column given the binary column. These statistics should either be known or estimated from a training set of typical images from the application area. Given models of this type, our task is to design an algorithm that will produce the sought-after binary and "fuzzified" columns in "real-time," that is as the noisy image arrives column by column.;In many applications signals are transformed into images by taking the squared norm of the Short-Time Fourier Transform of the signal, called the spectrogram. These spectrograms are extremely redundant; columns represent spectral information of overlapping time intervals. We, therefore, tested the possible applicability of image processing algorithms to the problem of increasing recognition in the hearing aid application. It was our hypothesis that for every spectrogram image of noisy speech there exist a binary image and its "fuzzified" image that (along with the noisy spectrogram image) contain the information essential for estimating the clean speech signal. The proposed processing was tested on 24 normal hearing subjects using the Modified Rhyme Test. Unfortunately, the specific choices that we made in the development of our processing methodology decreased the recognition of noisy speech signals. |
