Nonlinear feature extraction approaches with application to face recognition over large databases

The extraction of required features from the facial image is an important primitive task for face recognition. This paper evaluates different nonlinear feature extraction approaches, namely wavelet transform, radon transform and cellular neural networks (CNN). The scalability of the linear subspace techniques is limited as the computational load and memory requirements increase dramatically with the large database. In this work, the combination of radon and wavelet transform based approach is used to extract the multi-resolution features, which are invariant to facial expression and illumination conditions. The efficiency of the stated wavelet and radon based nonlinear approaches over the databases is demonstrated with the simulation results performed over the FERET database. This paper also presents the use of CNN in extracting the nonlinear facial features in improving the recognition rate as well as computational speed compared to other stated nonlinear approaches over the ORL database.

I. Introduction

The human skill of identifying thousands of people even after so many years excited many researches to focus on the face recognition system. The majority of the real world person identification and verification applications demand more robust, scalable and computationally efficient face recognition techniques under complex viewing and environmental conditions. Research concerning the face recognition started nearly in 1960's [1]. Different face recognition techniques have been proposed during last decades namely feature based, model based and appearance based techniques [2], [3]. In feature based techniques, the overall technique describes the position and size of each feature (eye, nose, mouth or face outline) [4]. In this approach, extracting features in different poses (viewing condition) and lighting conditions is a very complex task. For applications with large database, we have large set of features with different sizes and positions, making it difficult to identify the required feature points [4]. In the model based approach, a 3D model is constructed based on the facial variations in the image or important information related to the image. The difficulties in this approach are, we need high expensive camera (Stereo vision) to capture the facial variations clearly; further construction of 3D model is difficult and it takes more time to construct the model for large database [3]. The availability of large 3D data is also one of the complex task making the model based methods not suitable for real world application dealing with large databases.