Image-processing techniques for robust face recognition

Abstract

Current face-recognition algorithms can achieve a highly accurate performance under controlled conditions, such as unchanged light sources, frontal-view pose, no occlusion, neutral facial expression, etc. Face recognition has a wide range of applications, however it still has many technical and challenging issues to be solved, in particular when the faces under consideration have a very low resolution, different illumination conditions, arbitrary poses and are under occlusions. In order to achieve robust face recognition, we have investigated efficient techniques to solve some typical challenging problems for robust face recognition. First, among the facial features, the eyes play the most important role in face recognition and face hallucination. Most of the existing facial-feature detection and localization algorithms cannot work accurately when the faces are rotated or under poor lighting conditions. Therefore, in this research, an efficient algorithm for eye detection in face images is proposed. As the eye region always has the most variations in a face image, our algorithm uses a wavelet-based salient map, which can detect and reflect the most visually meaningful regions for coarse eye detection. With the aid of a pose-adapted eye template -which can handle eye regions with large rotation and pose variations, accurate eye positions can be localized. Furthermore, the position of the nose and mouth can be determined by considering both the saliency values in the salient map and the detected eye positions as geometric references. Second, face images under different illuminations represent a challenge for face recognition. In this research, we have discussed an efficient scheme for illumination compensation and the enhancement of face images. Our illumination model is universal; it does not require the assumption of a single-point light source. Thus, it circumvents and overcomes the limitations of the Lambertian model. The proposed approach can learn the average representations of face images under changing illuminations so as to compensate or enhance the face images, and also to eliminate the effect of different and uneven illuminations, while retaining the intrinsic properties of the face surface. Our experiments have provided promising results, demonstrating that our proposed methods are effective. Third, in order to achieve robust face recognition and to make face-recognition systems capable of identifying people at very low resolution, super-resolution (SR) technology is investigated. In this thesis, we first introduce facial-image super-resolution, which is also called face hallucination. In this research, an efficient mapping model is first proposed for face hallucination. Since we can observe and prove that the singular values of an image at one resolution, represented by singular value decomposition (SVD), have approximately linear relationships with their counterparts at other resolutions, the estimation of the corresponding singular values of the high-resolution (HR) face images becomes more reliable. From the signal-processing viewpoint, this can effectively preserve and reconstruct the dominant information in the reconstructed HR face images. The mapping scheme can be viewed as a "coarse-to-fine" estimation of HR face images. Compared to other, state-of-the-art algorithms, experiments have shown that our proposed face-hallucination scheme is practicable and effective. Fourth, a framework based on singular value decomposition (SVD) for performing both face hallucination and recognition simultaneously is also proposed. Conventionally, low-resolution (LR) face recognition is carried out by super-resolving the LR input face first, and then performing face recognition to identify the input face. By considering face hallucination and recognition simultaneously, the accuracy of both hallucination and recognition can be improved. In our algorithm, each face image is represented by using SVD. For each LR input face, the corresponding LR and high-resolution (HR) face-image pairs can then be selected from the face gallery. With the aid of face recognition, using the selected LR-HR pairs, the estimation of the mapping functions for interpolating the two matrices in the SVD representation of the corresponding HR face image can be more accurate. All these techniques can be integrated with both existing and new face recognition algorithms so as to achieve a robust and good performance level.