Learning Algorithms for Audio and Video Processing Independent Component Analysis and Support Vector Machine Based Approaches

Title of Dissertation LEARNING ALGORITHMS FOR AUDIO AND VIDEO PROCESSING INDEPENDENT COMPONENT ANALYSIS AND SUPPORT VECTOR MACHINE BASED APPROACHES Yuan Qi Master of Science Dissertation directed by Professor Rama Chellappa Department of Electrical and Computer Engineering In this thesis we propose two new machine learning schemes a Subband based Independent Component Analysis scheme and a hybrid Independent Component Analysis Support Vector Machine scheme and apply them to the problems of blind acoustic signal separation and face detection Based on a linear model classical Independent Component Analysis ICA provides a method of representing data as independent components In contrast to Principal Component Analysis PCA which decorrelates the data based on its covariance matrix ICA uses higher order statistics of the data to minimize the dependence between the components of the system output An important application of ICA is blind source separation However classical ICA algorithms do not work well for separation in the presence of noise or when performed on line Inspired by the psychoacoustic discovery that humans perceive and process acoustic signals in di erent frequency bands independently we propose a new algorithm subband based ICA that integrates ICA with time frequency analysis to separate mixed signals In subband based ICA the separations are performed in parallel in several frequency bands Wavelet decomposition and best basis selection in wavelet DCT packets can be incorporated into this algorithm Subband based ICA is computationally fast robust to noise and works well in an on line version when other ICA algorithms fail The virtually increased signal to noise ratio in those frequency bands where the separations are actually performed and the fact that subband signals i e wavelet coe cients are more peaky and heavy tailed distributed than the original signals both contribute to the success of subband based ICA Experimental results on separating noisy speech mixtures and musical signal mixtures demonstrate its e ectiveness In addition to separating mixed signals ICA can also be used as a feature extractor As argued by many researchers in the neural research area a principle of sensory information processing in the brain is redundancy reduction The ICA representation of the data follows this principle Also from a signal processing viewpoint ICA provides a nice way to cluster independent signals and hence leads to a better representation of signals than PCA Motivated by the feature extraction capability of ICA we propose a new hybrid unsupervised supervised learning scheme that integrates Independent Component Analysis with the Support Vector Machine SVM approach and apply this new learning scheme to the face detection problem SVM is a new powerful machine learning algorithmwhich is rooted in statistical learning theory As an approximate implementation of the Structural Risk Minimization SRM Principle proposed in statistical learning theory SVM tends to have good generalization performance One common characteristic shared by ICA and SVM is sparsity The ICA out put is sparse and the support vectors whose linear combination comprises the trained SVM are also sparse Thus integrating ICA with SVM yields a new hybrid hierarchical sparse learning scheme Speci cally for the face detection problem we use ICA in two di erent ways to extract low level features from a sliding window over an image and then apply SVM at a high level to classify the extracted ICA features as a face or not Ex perimental results show that using the rst method to extract ICA features and applying SVM for classi cation e ectively improves the detection system perfor mance compared with applying SVM directly to the original image data Finally as a general learning scheme hybrid ICA SVM can be applied to other pattern recognition problems as well as to face detection LEARNING ALGORITHMS FOR AUDIO AND VIDEO PROCESSING INDEPENDENT COMPONENT ANALYSIS AND SUPPORT VECTOR MACHINE BASED APPROACHES