Τεχνικές ψηφιακής επεξεργασίας εικόνας με εφαρμογές στις ανθρωποκεντρικές διεπαφές και την ανάλυση βίντεο

Abstract

The main research focus of this dissertation pertains to developing novel image processing and artificial intelligence techniques in order to perform face recognition, face verification, facial expression classification, and scene change detection in video sequences. Initially, a face recognition method is proposed, which aims to determine the identity of a test face by solving a number of simplified classification problems. This is achieved by applying dynamic training to a multi-step clustering scheme, during which discriminant clusters are created. The dynamic training allows creating a set of test-face-specific subspaces, which improves performance. Next, a face verification method is proposed. It applies a piecewise linear discriminant analysis process on similarity feature vectors and properly weighs the resulting discriminant hyper-planes. This method curbs problems related to insufficient numbers of training data and its expected classification performance is investigated via a series of simulations. Then, proper combinations of the various similarity scores are determined and used to produce a more accurate decision. In addition, a facial expression classification method is proposed. It operates by solving a number of two-class problems, a choice that is analytically justified. Then, it makes use of a class separability measure and applies an iterative process in order to extract high-quality features. Next, a classification scheme is developed where, during each of its steps, the most reliable classifier is identified and used in order to produce an more accurate decision. Lastly, a scene change detection method is proposed. It produces a set of enhanced eigen-audioframes that relate to a subspace where background noise changes are easily detected. Moreover, visual information is employed in order to align the audio scene change indications to neighboring video shot changes. In addition, certain video special-effects that are commonly used during scene changes are detected. By using various acoustic features to verify the scene change indications, the accuracy of this method increases further.