Fuyan Zhang

Recognizing partially occluded, expression variant faces from single training image per person with SOM and soft k-NN ensemble

Most classical template-based frontal face recognition techniques assume that multiple images per person are available for training, while in many real-world applications only one training image per person is available and the test images may be partially occluded or may vary in expressions. This paper addresses those problems by extending a previous local probabilistic approach presented by Martinez, using the self-organizing map (SOM) instead of a mixture of Gaussians to learn the subspace that represented each individual. Based on the localization of the training images, two strategies of learning the SOM topological space are proposed, namely to train a single SOM map for all the samples and to train a separate SOM map for each class, respectively. A soft k nearest neighbor (soft k-NN) ensemble method, which can effectively exploit the outputs of the SOM topological space, is also proposed to identify the unlabeled subjects. Experiments show that the proposed method exhibits high robust performance against the partial occlusions and variant expressions.

Constrained sampling for image retargeting

In this paper, we present a new approach for retargeting large images to mobile devices with small screens. As the core of image retargeting, information fidelity is adequately considered in terms of reservations of salient regions, edge integrity, and image layout. By taking these aspects as constraints, image retargeting is formulated as a constrained sampling task. Each pixel in image is first represented with a vector encoding the constraints. Then, pixels with the same vector values combine to form blocks, and the original image is thus converted into a graph representation. Thereafter, the sampling ratio of each block is determined with a balanced minimum cost flow algorithm. Final result is generated by an interpolated sampling scheme and direct scaling. Experiments demonstrate the effectiveness of the proposed approach.

Image completion based on views of large displacement

This paper presents an algorithm for image completion based on the views of large displacement. A distinct from most existing image completion methods, which exploit only the target image’s own information to complete the damaged regions, our algorithm makes full use of a large displacement view (LDV) of the same scene, which introduces enough information to resolve the original ill-posed problem. To eliminate any perspective distortion during the warping of the LDV image, we first decompose the target image and the LDV one into several corresponding planar scene regions (PSRs) and transform the candidate PSRs on the LDV image onto their correspondences on the target image. Then using the transformed PSRs, we develop a new image repairing algorithm, coupled with graph cut based image stitching, texture synthesis based image inpainting, and image fusion based hole filling, to complete the missing regions seamlessly. Finally, the ghost effect between the repaired region and its surroundings is eliminated by Poisson image blending. Our algorithm effectively preserves the structure information on the missing area of the target image and produces a repaired result comparable to its original appearance. Experiments show the effectiveness of our method.

Robust face recognition from a single training image per person with Kernel-based SOM-face

In this paper, a kernel-based SOM-face method is proposed to recognize expression variant faces under the situation of only one training image per person. Based on the localization of the face, an unsupervised kernel-SOM learning procedure is carried out to capture the common local features and the non-Euclidean structure of the image data, so that a compact and robust representation of the face can be obtained. Experiments on the FERET face database show that the Kernel-based SOM-face method can obtain higher recognition performance than the regular SOM-face method.

On volume distribution features based 3d model retrieval

In this paper, a 3D mesh retrieval method is proposed based on extracting geometric features of models. The method first finds three principal directions for a model by employing the principal component analysis method, and rotates the model to align it in a reference frame. Then, three sets of planes are used to slice the model along to the directions respectively. Subsequently, three character curves of the model can be obtained and be used as descriptor to key the model in 3D mesh model library. By comparing descriptors of two models, our method can compute similarity of models. Experiences show that our method is rapid, stable and robust to deal with various mesh models with arbitrary geometric and topological complexity.

Feature selection for high dimensional face image using self-organizing maps

While feature selection is very difficult for high dimensional, unstructured data such as face image, it may be much easier to do if the data can be faithfully transformed into lower dimensional space. In this paper, a new method is proposed to transform the high dimensional face images into low-dimensional SOM topological space, and then identify important local features of face images for face recognition automatically using simple statistics computed from the class distribution of the face image data. The effectiveness of the proposed method are demonstrated by the experiments on AR face databases, which reveal that up to 80% local features can be pruned with only slightly loss of the classification accuracy.

Smooth Approximation to Surface Meshes of Arbitrary Topology with Locally Blended Radial Basis Functions

AbstractIn this paper, we present a new approach for smooth approximation of surface meshes with arbitrary topology and geometry. The approach is based on the well-known radial basis functions (RBFs) for local shape approximation combined with a blending operator and a family of normalized weight functions for global surface construction. Our method first defines a local approximation using locally supported RBF for every vertex of the input mesh. A single global smooth surface is then constructed by blending the local approximations using weight functions associated with mesh vertices. A projection procedure is employed for visualizing the global surface using the local parameterization defined by barycentric coordinates for each facet of the input mesh. The approach provides a robust and efficient solution for smooth surface construction from various 3D mesh models.

An adaptive and efficient algorithm for polygonization of implicit surfaces

This paper describes an adaptive and efficient algorithm for polygonization of implicit surfaces, which consists of two steps: initial polygonization and adaptive refinement. The algorithm first generates an initial coarse triangular mesh from implicit surface using a variation of the traditional Marching Cubes (MC) Algorithm. And then the triangles in the coarse mesh are iteratively subdivided by employing a sampling rate that varies spatially according to local complexity of the surface. The new created vertices in refined mesh are projected onto the implicit surface by gradient descent method. Consequently, the algorithm produces the minimum number of polygons required to approximate the surface with a desired precision and the final mesh is simplicial complex. Our algorithm can be used in the real-time environment of visualization of implicit surfaces.

Weighted SOM-Face: selecting local features for recognition from individual face image

In human face recognition, different facial regions have different degrees of importance, and exploiting such information would hopefully improve the accuracy of the recognition system. A novel method is therefore proposed in this paper to automatically select the facial regions that are important for recognition. Unlike most of previous attempts, the selection is based on the facial appearance of individual subjects, rather than the appearance of all subjects. Hence the recognition process is class-specific. Experiments on the FERET face database show that the proposed methods can automatically and correctly identify those supposed important local features for recognition and thus are much beneficial to improve the recognition accuracy of the recognition system even under the condition of only one single training sample per person.

Binary Alpha-Plane Assisted Fast Motion Estimation of Video Objects in Wavelet Domain

In this paper, we present a novel approach to motion estimation (ME) of arbitrarily shaped video objects in wavelet domain. We explore the guiding role of binary alpha-plane in assisting ME of video objects and first devise a new block matching scheme of alpha-plane, by exploiting boundary expansion and boundary masks. To eliminate shift-variance, we modify low-band-shift (LBS) method via substituting variable-size block for wavelet block. Combining the modified LBS with a hierarchical structure, we further present a multiscale ME approach. Extensive experiments show that the proposed approach outperforms most of previous methods in terms of both subjective quality and objective quality. Moreover, significant reduction is achieved in computational complexity (89.05% at most) and memory requirement.