Date of Award
Doctor of Philosophy (PhD)
Richard Yehwhei Han
Pei Hsui Yeh
Multimedia data is increasingly important in scientific discovery and people's daily lives. Content of massive multimedia is often diverse and noisy, and motion between frames is sometimes crucial in analyzing those data. Among all, still images and videos are commonly used formats. Images are compact in size but do not contain motion information. Videos record motion but are sometimes too big to be analyzed. Sequential images, which are a set of continuous images with low frame rate, stand out because they are smaller than videos and still maintain motion information. This thesis investigates features in different types of noisy sequential images, and the proposed solutions that intelligently combined multiple features to successfully retrieve visual information from on-line videos and cloudy satellite images. The first task is detecting supraglacial lakes above ice sheet in sequential satellite images. The dynamics of supraglacial lakes on the Greenland ice sheet deeply affect glacier movement, which is directly related to sea level rise and global environment change. Detecting lakes above ice is suffering from diverse image qualities and unexpected clouds. A new method is proposed to efficiently extract prominent lake candidates with irregular shapes, heterogeneous backgrounds, and in cloudy images. The proposed system fully automatize the procedure that track lakes with high accuracy. We further cooperated with geoscientists to examine the tracked lakes and found new scientific findings. The second one is detecting obscene content in on-line video chat services, such as Chatroulette, that randomly match pairs of users in video chat sessions. A big problem encountered in such systems is the presence of flashers and obscene content. Because of various obscene content and unstable qualities of videos capture by home web-camera, detecting misbehaving users is a highly challenging task. We propose SafeVchat, which is the first solution that achieves satisfactory detection rate by using facial features and skin color model. To harness all the features in the scene, we further developed another system using multiple types of local descriptors along with Bag-of-Visual Word framework. In addition, an investigation of new contour feature in detecting obscene content is presented.
Liang, Yu-Li, "Feature Extraction in Sequential Multimedia Images: with Applications in Satellite Images and On-line Videos" (2012). Computer Science Graduate Theses & Dissertations. 61.