| Summary: | Face processing provides a user-friendly and nonintrusive means of human identification. The fundamental problems in face identification are uncontrolled illumination conditions and high degree of features complexity. Many studies focus on illumination normalization and dimensionality reduction. Yet, simplistic and realistic criteria are demanded. In this thesis, a framework that enhances feature selection for simultaneous detection and recognition of face using Hough Transform (HT) is introduced. Subsequently, Anisotropic Diffusion Illumination Normalization technique based DCT (ASDCT) is proposed to compensate illumination and improve the decorrelation of features. Meanwhile, an algorithm named' Integrated Histograms based on Discrete Cosine Transform' (HIDCT) is developed to increase data compaction and reduce redundancy. Reduced number of features is obtained through rotating histogram plots by 30 degrees. Then HT matrices are extracted for face recognition. The best illumination normalization technique is known by comparing 22 illumination normalization techniques in the literature. Then 19 coefficients with largest magnitude are selected. Discrete Cosine Transform (DCT) is exploited on the reduced features to form new coefficients that optimize lossy compression techniques. The performance of the system is tested using standard research databases. Template matching and nearest neighbor classifiers are used for similarity measurements. Experimental results indicated that, the enhanced framework is very effective and accomplished an average detection time of 1.199 seconds, and minimum recognition accuracy of 95.625%. ASDCT technique showed up to 97.700% verification rate at 1.0% of False Acceptance Rate (FAR). This supersedes Appearance based techniques such as Linear Discriminant Analysis (LDA), and Kernel Principal Component Analysis (KPCA). The level of compression of HIDCT algorithm is at least three times the conventional DCT. The proposed HIDCT is simple to implement and can accommodate large users within a small memory space. Furthermore, the three frameworks encourage real world applications for efficient identifications.
|
|---|