A contribution for single and multiple faces recognition using feature-based approaches

Abstract

Among biometric recognition systems, face biometrics plays an important role in research activities and security applications since face images can be acquired without any knowledge of individuals. Nowadays a huge amount of digital images and video sequences have been acquired mainly from uncontrolled conditions, frequently including noise, blur, occlusion and variation on scale and illumination. Because of these issues, face recognition (FR) is still an active research area and becomes a complex problem and a challenging task. In this context, the motivation comes from the fact that recognition of faces in digital images with complex background and databases of face images have become one of the successful applications of Computer Vision. Hence, the main goal of this work is to recognize one or more faces from still images with multiple faces and from a database of single faces obtained under different conditions. To work with multiple face images under varying conditions, a semi-supervised approach proposed based on the invariant and discriminative power of local features. The extraction of local features is done using Speeded-Up Robust Features (SURF). The search for regions from which optimal features can be extracted is fulfilled by an improved ABC algorithm. To fully exploit the proposed approach, an extensive experimental analysis was performed. Results show that this approach is robust and efficient for face recognition applications except for faces with non-uniform illumination. In the literature, a significant number of single FR researches are based on extraction of only one feature and machine learning approaches. Besides, existing feature extraction approaches broadly use either global or local features. To obtain relevant and complementary features from face images, a face recognition methodology should consider heterogeneous features and semi-global features. Therefore, a novel hierarchical semi-supervised FR approach is proposed based on extraction of global, semi-global and local features. Global and semi-global features are extracted using Color Angles (CA) and edge histogram descriptors (EHD) meanwhile only local features are extracted using SURF. An extensive experimental analysis using the three feature extraction methods was done first individually followed by a three-stage hierarchical scheme using the face images obtained under two different lighting conditions with facial expression and slight scale variation. Furthermore, the performance of the approach was also analyzed using global, semi-global and local features combinations for CA and EHD. The proposed approach achieves high recognition rates considering all image conditions tested in this work. In addition to this, the results emphasize the influence of local and semi-global features in the recognition performance. In both, single face and multiple faces approaches, the main achievement is the high performance obtained only from the discriminative capacity of extracted features without any training schemes.