Date of Award
Doctor of Philosophy in Computing Sciences - (Ph.D.)
James A. McHugh
Local Binary Patterns (LBP), which is one of the well-known texture descriptors, has broad applications in pattern recognition and computer vision. The attractive properties of LBP are its tolerance to illumination variations and its computational simplicity. However, LBP only compares a pixel with those in its own neighborhood and encodes little information about the relationship of the local texture with the features. This dissertation introduces a new Feature Local Binary Patterns (FLBP) texture descriptor that can compare a pixel with those in its own neighborhood as well as in other neighborhoods and encodes the information of both local texture and features. The features encoded in FLBP are broadly defined, such as edges, Gabor wavelet features, and color features. Specifically, a binary image is first derived by extracting feature pixels from a given image, and then a distance vector field is obtained by computing the distance vector between each pixel and its nearest feature pixel defined in the binary image. Based on the distance vector field and the FLBP parameters, the FLBP representation of the given image is derived. The feasibility of the proposed FLBP is demonstrated on eye detection using the BioID and the FERET databases. Experimental results show that the FLBP method significantly improves upon the LBP method in terms of both the eye detection rate and the eye center localization accuracy.
As LBP is sensitive to noise especially in near-uniform image regions, Local Ternary Patterns (LTP) was proposed to address this problem by extending LBP to three-valued codes. However, further research reveals that both LTP and LBP achieve similar results for face and facial expression recognition, while LTP has a higher computational cost than LBP. To improve upon LTP, this dissertation introduces another new local texture descriptor: Local Quaternary Patterns (LQP) and its extension, Feature Local Quaternary Patterns (FLQP). LQP encodes four relationships of local texture, and therefore, it includes more information of local texture than the LBP and the LTP. FLQP, which encodes both local and feature information, is expected to perform even better than LQP for texture description and pattern analysis. The LQP and FLQP are applied to eye detection on the BioID database. Experimental results show that both FLQP and LQP achieve better eye detection performance than FLTP, LTP, FLBP and LBP. The FLQP method achieves the highest eye detection rate.
Gu, Jiayu, "Innovative local texture descriptors with application to eye detection" (2014). Dissertations. 143.