Electrocardiogram(EKG)

    Non-invasive wireless fetal heart rate monitoring during pregnancy

    Heart rate variability
       
Advanced Digital Signal Processing

    Filter Banks and Sub-band Coding Design for Images

    We explored different design methods for different Perfect Reconstruction Filter Banks.First, power symmetric FIR filter banks are designed to meet some constraints, then quantization/compression is studied to minimize bits allocation while retaining good reconstruction. Next effectiveness of Subband coding is put into test with lossy image compression through 4 different filter bank design problems:1- a PR FIR QMF bank, 2- a Linear uniform DFT, 3- a non-uniform three channel PR, and 4- a 2-D PR FIR bank.     


    Signal modeling, spectrum estimation, and adaptive filtering

    We explore the efficient Levinson-Durbin recursive algorithm, then use it in the solution of a signal modeling problem. We look into adaptive filtering for noise cancellation when stationarity is not present. Two methods are explored: Fast least mean square method and the recursive least square method.



    Filter Bank Design and Sub-band coding for Digital Audio

    A nine-octave filter bank was developed using an order 64 linear phase FIR filter with steep attenuation outside the passband as a prototype. A simple, yet effective quantization scheme was used to compress the sub-band data. An output that is perceptually almost identical to the original can be achieved with at least 6-fold compression. This scheme is compared to the filter bank used in MPEG-1 and is shown to perform better, at least using this quantization scheme. When compared to a fully-fledged MPEG-1, layer 3 codec, the layer 3 codec achieved better compression, but that is to be expected as it uses a much more involved quantization scheme as well as source coding. These results show that the nine-octave filter bank is very suitable for audio compression.


    Linear Prediction, Synthesis and Spectrum Estimation

    We use linear prediction to compress speech. Linear prediction models the vocal tract using an all pole model that changes regularly. This model is based on the physics of human speech production and manages to remove a lot of the redundancy in the speech signal. The remaining signal must then be coded in some way and we present two methods that give varying quality but also very different bit-rates. The choice between these two methods depends on the bit-rate requirements and the environment the codec is to be used in.


Speech and Audio Processing

    Sonification of visual information

    We discuss methods used to sonify image and video data. For a varietyof reasons, sonication can help users acquire more a comprehensive understanding
about visual information. Using Matlab, we attempt to map visual information to auditory information by applying traditional digital image and video processing techniques to both still images and video samples. After applying different techniques to three video samples, we combine the approaches to construct one coherent sonification system.


    Emotional Prosody detection from speech information

    We used an in-house auditory model that predicted better classification of 4 types of emotions compared to other training methods such as neural networks and Support Vector Machine.

    Biology of Hearing Grant Proposal

    Perception is the ability to identify, interpret, and attach meaning to sound. Our main goal of this study is to answer questions about such perception by measuring Steady State Response (SSR) using simple Sinusoidally Amplitude Modulated (SAM) pure tones. We will record human brain responses using Magnetoencephalography (MEG) that is ideal for auditory cortex recording with high temporal resolution.


    Alien Auditory system proposal                                

    If we look at audiogram for terrestrial creatures, there is no one terrestrial single creature that exhibit a wide frequency range of hearing that span 20 – 100,000 Hz (12.3 Octave). Besides, although extensively studied, the auditory periphery is still not fully understood as to what determines frequency selectivity, absolute threshold, and so on. As a result, we adopt a design mechanism of selecting 2 audiograms from the more understood terrestrial audiograms as draft model for our design. One should be more in the low to mid frequency and the second from the mid to high frequency. We select human as the species of our extraterrestrial live, with a modified pair of ears to fit the requirement of high absolute hearing sensitivity in the range of 20 to
2,000Hz. On the other hand, we pick Horseshoe bat as the second draft model to use to satisfy the high frequency requirement. Hence it is modified accordingly. Our creature, will have 4 ears. Two are placed same as for humans, while the second pair of ears sit on top of the head toward the frontal lobe, and shares lots of properties of bat ears except for the modification suggested below. We name our design Humechon. Both pair of ears, corresponding to two complementary auditory systems with some overlap, will share the same higher brain functions. Each system will have separate outer, middle, and inner ear pair.


Statistical and Neural Pattern Recognition

    Face Recognition   

    We explored holistic approaches for face recognition,and then proposed a new method driven by our neuroscience understanding of visual
system. First, we built a principal component analysis (PCA) and a linear discriminant analysis (LDA) named as Eigenface and Fisherface respectively. Then we modified PCA algorithm to work for two-dimensional images without the requirements of vectorizing images. Then a kernel based PCA is explored that exploit high order statistics for recognition. Finally, we proposed an adaptive filter in the frequency domain that correlates a sequence of test and training images of same subject. All four methods were tested on three different databases. High detection rate was possible with higher number of training images per subject; we got close to 100% with reasonable ratio of training to test images.

    Gender Recognition 

    We applied support vector machine to 2 classification problems: face detection, and gender classification. For gender classification, SVM achieved superior results with linear kernel applied to intensities of images. Other methods that reduce the dimensionality of our data, before applying SVM, were compared from ICA, PCA, and moments, but each suffered from some weaknesses delivering less superior results. Linear kernel outperformed others for our dataset, although many experiments from the literature favor RBF kernel. For face detection problem, we were able to improve on our previous results of using Eigenface but not for Fisherface. Few percents of detection improvements overall with a large multi-class problem is good sign in of power of SVM.