英文摘要 |
Face analysis is a popular biometric technique and has many potential applications. In the constrained situation, face analysis has achieved excellent performance. However, in the unconstrained situation, the performance is still far behind satisfactory due to various factors such as illumination, pose, expression and low-resolution. In this thesis, we study some unconstrained face analysis tasks and propose our improved algorithms. The main contributions include the following issues:
1.To deal with the video based face recognition, we propose a joint space learning method to simultaneously identify the most representative samples and discriminative feature from face video. Joint space learning is formulated as a matrix minimization problem with respect to both the columns (samples) and rows (features). Then an alternate minimization algorithm is developed to monotonically decrease the joint loss function. In addition, randomized techniques are applied to capture the nonlinear structure in unconstrained data, so that both accuracy and efficiency can be improved.
2.To improve the efficiency of video based face recognition, we propose to learn discriminative and compact binary codes for image set. To do this, we propose to embed the Hadamard code into the hashing function. This process not only leverages discriminative information but favors an information-theoretic criterion to yield high-quality codes. The low rank constraint is introduced to reduce the redundance of the image set. Moreover, we use a anchor points based kernel method to further improve the performance of the algorithm.
3.To deal with the unconstrained gender recognition, we propose to learn multiple order local binary patterns as feature descriptor. Specifically, we extract features according to three different statistical methods, i.e., single pixel value, mean and variance.
Then, we further develop a localized multi-boost learning algorithm to combine these features for classification. Experiments show that the proposed method can effectively reduce the influence of the unconstrained factors.
4.To deal with the unconstrained multi-view clustering problem, we propose a new Dictionary learning framework, called Nonnegative Dictionary Pair Learning, for robust multi-view clustering. To do this, we propose to learn a semantic projection and a feature projection jointly. A consistency constraint and a local geometric preserving constraint are combined to push the clustering solution in each view towards a common consensus. Then an alternate minimization algorithm called proximal alternating linearized minimization algorithm (PALM) is developed to monotonically decrease the joint loss function.
In summary, in this thesis, we systematically study some unconstrained face analysis problems like identification recognition, image set hashing, gender recognition and multi-view clustering. Our proposed works improve the performance of the related challenges. |
修改评论