Catálogo de publicaciones - libros

In this paper, we propose an Active Appearance Models (AAMs) fitting algorithm, adaptive fitting algorithm, to localize an object in an image containing occlusion. The adaptive fitting algorithm conducts the fitting problem of AAMs containing object occlusion in a statistical framework. We assume that the residual errors can be treated as mixture statistical model of Gaussian and uniform model. We then reformulated the basic fitting algorithm and maximum a-posteriori (MAP) estimation algorithm of model parameter for AAMs to make the adaptive fitting algorithm. Extensive experiments are provided to demonstrate our algorithm.

Traditional personal authentication methods have many instinctive defects. Biometrics is an effective technology to overcome these defects. Among the available biometric approaches, iris recognition is one of the most accurate techniques. Combining the left and the right irises of same persons can improve the authentication accuracy and reduce the spoof attack risks. Furthermore, the fusion need not add any other hardware to the existing iris recognition systems. This paper investigates the feasibility of fusing both irises for personal authentication and the performance of some very simple fusion strategies. The experimental results show that the difference between the left and the right irises of the same persons is close to the difference between the irises captured from different persons. And combining the information of both irises can dramatically improve the authentication accuracy even when the quality of the iris images are not good enough. The results also show that the Minimum and the Product strategies can obtain the perfect performance, i.e. both FARs and FRRs of these two strategies can be reduce to 0%.

Recognizing humans, estimating their pose and segmenting their body parts are key to high-level image understanding. Because humans are highly articulated, the range of deformations they undergo makes this task extremely challenging. Previous methods have focused largely on heuristics or pairwise part models in approaching this problem. We propose a bottom-up , similar to parsing, of increasingly more complete partial body masks guided by a set of parse rules. At each level of the growing process, we evaluate the partial body masks directly via shape matching with exemplars (and also image features), without regard to how the hypotheses are formed. The body is evaluated as a whole, not the sum of its parts, unlike previous approaches. Multiple image segmentations are included at each of the levels of the growing/parsing, to augment existing hypotheses or to introduce ones. Our method yields both a pose estimate as well as a segmentation of the human. We demonstrate competitive results on this challenging task with relatively few training examples on a dataset of baseball players with wide pose variation. Our method is comparatively simple and could be easily extended to other objects. We also give a learning framework for parse ranking that allows us to keep fewer parses for similar performance.

This paper presents a large scale general purpose image database with human annotated ground truth. Firstly, an all-in-all labeling framework is proposed to group visual knowledge of three levels: scene level (global geometric description), object level (segmentation, sketch representation, hierarchical decomposition), and low-mid level (2.1D layered representation, object boundary attributes, curve completion, etc.). Much of this data has not appeared in previous databases. In addition, is used to organize visual elements to facilitate top-down labeling. An annotation tool is developed to realize and integrate all tasks. With this tool, we’ve been able to create a database consisting of more than 636,748 annotated images and video frames. Lastly, the data is organized into 13 common subsets to serve as benchmarks for diverse evaluation endeavors.

In this paper, we present an automatic human portrait system based on the And-Or graph representation. The system can automatically generate a set of life-like portraits in different styles from a frontal face image. The system includes three subsystems, each of which models hair, face and collar respectively. The face subsystem can be further decomposed into face components: eyebrows, eyes, nose, mouth, and face contour. Each component has a number of distinct sub-templates as a leaf-node in the And-Or graph for portrait. The And-Or graph for portrait is like a ”mother template” which produces a large set of valid portrait configurations, which is a ”composite templates” made of a set of sub-templates. Our approach has three novel aspects:(1) we present an And-Or graph for portrait that explains the hierarchical structure and variability of portrait and apply it into practice; (2) we combine hair, face and collar into a system that solves a practical problem; (3) The system can simultaneously generate a set of impressive portraits in different styles. Experimental results demonstrate the effectiveness and life-likeness of our approach.

In this paper, we present a framework for object categorization via sketch graphs, structures that incorporate shape and structure information. In this framework, we integrate the learnable And-Or graph model, a hierarchical structure that combines the reconfigurability of a stochastic context free grammar(SCFG) with the constraints of a Markov random field(MRF), and we sample object configurations as training templates from this generative model. Based on these synthesized templates, four steps of discriminative approaches are adopted for cascaded pruning, while a template matching method is developed for top-down verification. These synthesized templates are sampled from the whole configuration space following the maximum entropy constraints. In contrast to manually choosing data, they have a great ability to represent the variability of each object category. The generalizability and flexibility of our framework is illustrated on 20 categories of sketch-based objects under different scales.

This paper presents a Bayesian inference algorithm for image layer representation [26], 2.1D sketch [6], with mixed Markov random field. 2.1D sketch is an very important problem in low-middle level vision with a synthesis of two goals: segmentation and 2.5D sketch, in other words, it is to consider 2D segmentation by incorporating occulision/depth explicitly to get the partial order of final segmented regions and contour completion in the same layer. The inference is based on Swendsen-Wang Cut (SWC) algorithm [4] where there are two types of nodes, instead of all nodes being the same type in traditional MRF model, in the graph representation: atomic regions and their open bonds desribed by address variables. These makes the problem a mixed random field. Therefore, two kinds of energies should be simultaneously minimized by maximizing a joint posterior probability: one is for region coloring/layering, the other is for the assignments of address variables. Given an image, its primal sketch is computed firstly, then some atomic regions can be obtained by completing some sketches into a closed contour. At the same time, T-junctions are detected and broken into terminators as the open bonds of atomic regions after being assigned the ownership between them and atomic regions. With this graph representation, the presented inference algorithm is performed and satisfactory results are shown in the experiments.

A feature-based technique for separating specular and diffuse components of a single image is presented. In the proposed approach, Shafer’s dichromatic reflection model is utilized, which assumed a light reflected from a surface point is linearly composed of diffuse and specular reflections. The major idea behind the proposed method is to classify the boundary pixels of the input image to be specular-related or diffuse-related. A fuzzy integral process is proposed to classify boundary pixels based on their local evidences, including specular and diffuse estimation information. Based on the classification result of boundary pixels, an integration method is evoked to reconstruct the specular and diffuse components of the input image, respectively. Unlike previous researches, the proposed method has no color segmentation and iterative operations. The experimental results have demonstrated that the proposed method can perform dichromatic reflectance separation effectively with small misadjustments and rapid convergence.

Restoration and segmentation in corrupted text images are very important processing steps in digital image processing and several different methods were proposed in the open literature. In this paper, the restoration and segmentation problem in corrupted color text images are addressed by tensor voting and statistical method. In the proposed approach, we assume to have corruptions in text images. Our approach consists of two steps. The first one uses the tensor voting algorithm. It encodes every data point as a particle which sends out a vector field. This can be used to decompose the pointness, edgeness and surfaceness of the data points. And then noises in a corrupted region are removed and restored by generalized adaptive vector sigma filters iteratively. In the second step, density mode detection and segmentation using statistical method based on Gaussian mixture model are performed in values according to hue and intensity components in the image. The experimental results show that proposed approach is efficient and robust in terms of restoration and segmentation corrupted text images.

Moving cast shadow causes serious problem while segmenting and extracting foreground from image sequences, due to the misclassification of moving shadow as foreground. This paper proposes a boosting discriminative model for moving cast shadow detection. Firstly, color invariance subspace and texture invariance subspace are obtained by the color and texture difference between current image and background image; then, boosting is selected based on theses subspaces to discriminate cast shadow from moving objects; finally, temporal and spatial coherence of shadow and foreground is employed on Discriminative Random Fields for accurate image segmentation through graph cut. Results show that the proposed method excels classical method both in indoor and outdoor scene.