Modified least trimmed squares method for face recognition / Nur Azimah Abdul Rahim

Face recognition involves the comparison of a given face with other faces in a database. A fully automated face recognition system would consist of several subsystems including face detection, normalization and authentication. Features of the face to be normalized include the size, orientation and the illumination. Facial feature detection must first be performed before any of the face recognition methods can be applied. A good face detection system would take care of most of these processes. There exist various frameworks and algorithms for a face recognition system. However, most of these frameworks are only reliable when the face is captured under controlled environment. The face recognition method is very much affected by noise or occlusion, which can be seen as grain in film and pixel variations if in digital images and their presence caused varying intensity in the image pixels instead of true pixel values. For most face recognition algorithms, partial occlusions affect the performance of the algorithm. This research addressed severe contamination or occlusion presence in a face recognition based on image data. A modified version of the existing least trimmed square, LTS method with genetic algorithm (LTS with GAs) was proposed to cater the problem of noise or occlusion and improve the performance of face recognition. In the proposed algorithm, the contaminated observations are distinguished in C-steps as every observation will be assigned a weight based on a cutoff value which will give a zero weight for any observations with residual error greater than the cutoff value and a weight "one" (1) otherwise. A robust standard error was used in this research for a more precise cutoff value in determining outliers. Benchmark datasets, namely the AT&T and Yale which contain occluded query images were used to examine the performance of the proposed method. The query images were contaminated with salt and pepper noise and the recognition rates was measured when the contaminated images were used as a query image in the context of linear regression. The best method was the one being least affected by the occluded images and produces highest recognition rates. The proposed approach performs almost as good as the FAST-LTS method with highest recognition rate as compared to other methods for Yale dataset. A simulation study was also done to further assess the performance of the modified approach alongside with several LTS based methods for large data sets which were contaminated with different levels of noise. The genetic algorithm configuration for n (number of observations) and p (parameter) was changed to assess the performance of modified method. The proposed method does not lose its robustness property, and its estimates are still unbiased and have a minimum variance in this configuration. It can be concluded that the modified algorithm decreases the biases, the variances and the mean squared errors of the LTS estimators. This research contributes to method in face recognition, which can be used in broad fields such as video and image processing, human-computer interaction, criminal identification, homeland security and numerous consumer applications.