Date of Award:


Document Type:


Degree Name:

Doctor of Philosophy (PhD)


Computer Science


Xiaojun Qi


Face recognition under illumination is really challenging. This dissertation proposes four effective methods to produce illumination-invariant features for images with various lev- els of illuminations. The proposed methods are called logarithmic fractal dimension (LFD), eight local directional patterns (ELDP), adaptive homomorphic eight local directional pat- terns (AH-ELDP), and complete eight local directional patterns (CELDP), respectively.

LFD, employing the log function and the fractal analysis (FA), produces a logarithmic fractal dimension (LFD) image that is illumination-invariant. The proposed FA feature- based method is an effective edge enhancer technique to extract and enhance facial features such as eyes, eyebrows, nose, and mouth.

The proposed ELDP code scheme uses Kirsch compass masks to compute the edge responses of a pixel's neighborhood. It then uses all the directional numbers to produce an illumination-invariant image.

AH-ELDP first uses adaptive homomorphic filtering to reduce the influence of illumi- nation from an input face image. It then applies an interpolative enhancement function to stretch the filtered image. Finally, it produces eight directional edge images using Kirsch compass masks and uses all the directional information to create an illumination-insensitive representation.

CELDP seamlessly combines adaptive homomorphic filtering, simplified logarithmic fractal dimension, and complete eight local directional patterns to produce illumination- invariant representations.

Our extensive experiments on Yale B, extended Yale B, CMU-PIE, and AR face databases show the proposed methods outperform several state-of-the-art methods, when using one image per subject for training.

We also evaluate the ability of each method to verify and discriminate face images by plotting receiver operating characteristic (ROC) curves which plot true positive rates (TPR) against the false positive rates (FPR).

In addition, we conduct an experiment on the Honda UCSD video face database to simulate real face recognition systems which include face detection, landmark localization, face normalization, and face matching steps. This experiment, also, verifies that our proposed methods outperform other state-of-the-art methods.