Image super-resolution is a typically ill-conditioned (if not ill-posed) discrete inverse problem in which motion estimation, deblurring, denoising and scaling-up tasks are usually involved to obtain a higher optical resolution image from a sequence of low-resolution images. The theme of this thesis is the high quality facial image super-resolution reconstruction (SRR) from low quality ones, including SRR using single image, SRR from sequences and blind super-resolution (BSR). This thesis offers contributions in three main areas. First, we propose an adaptive manifold learning method for frontal facial image SRR from a single image. Single frame SRR is mathematically more ill-posed relative to multi-frame SRR. Since the high-frequency information of facial images is usually contained in the local subtle variations, we analyze the local structures of facial images on the sub-manifolds using a new manifold learning method, Locality Preserving Projections (LPP). We accomplish the adaptive sample selection by searching out patches online in the LPP sub-space, which makes the resultant training set tailored to the testing patch, and then effectively restore the lost high-frequency components of the low-resolution face image by patched-based eigen-transformation using the dynamic training set. Second, looking upon the patches of face image as some specific classes of signals, we propose a new regularization method for facial image SRR in which the orthogonal complement subspace of image patches is used as a regularization constraint and combined with the classic reconstruction constraint under a Bayesian framework. Under affine motion model, we derive a more general Jacobian of the warped image w.r.t motion parameters by extending the four-neighbor image interpolation to the estimation of image gradient field. We further develop three iterative algorithms for the joint estimation of motion parameters and HR image by means of full or partial differential expansions of the cost functional, which helps convert the original non-linear least square problem to a linear one. An adaptive method is also presented for the computation of regularization parameter. Third, using a sequence of LR images under affine motion, we propose a method for joint registration, non-parametric blur identification and high-resolution image estimation. We discuss the estimation method for blurs with arbitrary shape and size beginning with the proof of an equivalent of th...
修改评论