The wide availability of visual data via social media and the internet, coupled with the demands of the security community have led to an increased interest in visual recognition. Recent research has focused on improving the accuracy of recognition techniques in environments where variability is well controlled. However, applications such as identity verification often operate in unconstrained environments. Therefore there is a need for more robust recognition techniques that can operate on data with considerable noise.
Many statistical recognition techniques rely on principal component analysis (PCA). However, PCA suffers from the presence of outliers due to occlusions and noise often encountered in unconstrained settings. In this thesis we address this problem by using L1-PCA to minimize the effect of outliers in data. L1-PCA is applied to several statistical recognition techniques including eigenfaces and Grassmannian learning. Several popular face databases are used to show that L1-Grassmann manifolds not only outperform, but are also more robust to noise and occlusions than traditional L2-Grassmann manifolds for face and facial expression recognition. Additionally a high performance GPU implementation of L1-PCA is developed using CUDA that is several times faster than CPU implementations.
Library of Congress Subject Headings
Optical pattern recognition; Image processing--Digital techniques; Principal components analysis; CUDA (Computer architecture)
Computer Engineering (MS)
Department, Program, or Center
Computer Engineering (KGCOE)
Sonia Lopez Alarcon
Johnson, Matthew, "Robust Recognition using L1-Principal Component Analysis" (2015). Thesis. Rochester Institute of Technology. Accessed from
RIT – Main Campus