Problems in image matching, saliency detection in images, and background detection in video are studied. Algorithms based on approximate nearest-neighbor matching are proposed to solve problems in these related domains. Image patches are quantized into features using a special Walsh-Hadamard transform, and put into a propagation-assisted kd-tree for indexing and search. Image saliency and background-detection algorithms are then derived by looking at patch similarity over time and space.
Library of Congress Subject Headings
Nearest neighbor analysis (Statistics); Image processing--Digital techniques; Computer vision--Data processing; Optical pattern recognition
Department, Program, or Center
Computer Science (GCCIS)
Allmann, Josh, "Approximate nearest neighbors for recognition of foreground and background in images and video" (2013). Thesis. Rochester Institute of Technology. Accessed from
RIT – Main Campus