Catálogo de publicaciones - libros

We analyse a series of approaches to evolve images. It is motivated by combining Gaussian blurring, the Mean Curvature Motion (used for denoising and edge-preserving), and maximal blurring (used for inpainting). We investigate the generalised method using the combination of second order derivatives in terms of gauge coordinates.

For the qualitative behaviour, we derive a solution of the PDE series and mention its properties briefly. Relations with general diffusion equations are discussed. Quantitative results are obtained by a novel implementation whose stability and convergence is analysed.

The practical results are visualised on a real-life image, showing the expected qualitative behaviour. When a constraint is added that penalises the distance of the results to the input image, one can vary the desired amount of blurring and denoising.

The paper proposes an approach to superimpose virtual contents for advertising in an existing image sequence with no or minimal user interaction. Our approach automatically recognizes planar surfaces in the scene over which a billboard can be inserted for seamless display to the viewers. The planar surfaces are segmented in the image frame using a homography dependent scheme. In each of the segmented planar regions, a rectangle with the largest area is located to superimpose a billboard into the original image sequence. It can also provide a viewing index based on the occupancy of the virtual real estate for charging the advertiser.

Background subtraction is a method commonly used to segment objects of interest in image sequences. By comparing new frames to a background model, regions of interest can be found. To cope with highly dynamic and complex environments, a mixture of several models has been proposed. This paper proposes an update of the popular Mixture of Gaussian Models technique. Experimental analysis shows a lack of this technique to cope with quick illumination changes. A different matching mechanism is proposed to improve the general robustness and a comparison with related work is given. Finally, experimental results are presented to show the gain of the updated technique, according to the standard scheme and the related techniques.

Many existing tone reproduction schemes are based on the use of a single high dynamic range (HDR) image and are therefore unable to accurately recover the local details and colors of the scene due to the limited information available. Accordingly, the current study develops a novel tone reproduction system which utilizes two images with different exposures to capture both the local details and color information of the low- and high-luminance regions of a scene. By computing the local region of each pixel, whose radius is determined via an iterative morphological erosion process, the proposed system implements a pixel-wise local tone mapping module which compresses the luminance range and enhances the local contrast in the low-exposure image. And a local color mapping module is applied to capture the precise color information from the high-exposure image. Subsequently, a fusion process is then performed to fuse the local tone mapping and color mapping results to generate highly realistic reproductions of HDR scenes.

This paper presents a novel, automated method to remove partial occlusion from a single image. In particular, we are concerned with occlusions resulting from objects that fall on or near the lens during exposure. For each such foreground object, we segment the completely occluded region using a geometric flow. We then look outward from the region of complete occlusion at the segmentation boundary to estimate the width of the partially occluded region. Once the area of complete occlusion and width of the partially occluded region are known, the contribution of the foreground object can be removed. We present experimental results which demonstrate the ability of this method to remove partial occlusion with minimal user interaction. The result is an image with improved visibility in partially occluded regions, which may convey important information or simply improve the image’s aesthetics.

High-capacity image watermarking scheme aims at maximize bit rate of hiding information, neither eliciting perceptible image distortion nor facilitating special watermark attack. Texture, in preattentive vision, delivers itself by concise high-order statistics, and holds high capacity for watermark. However, traditional distortion constraint, e.g. just-noticeable-distortion (JND), cannot evaluate texture distortion in visual perception and thus imposes too strict constraint. Inspired by recent work of image representation [9], which suggests texture extraction and mix probability principal component analysis for learning texture feature, we propose a distortion measure in the subspace spanned by texture principal components, and an adaptive distortion constraint depending on image local roughness. The proposed spread-spectrum watermarking scheme generates watermarked images with larger SNR than JND-based schemes at the same distortion level allowed, and its watermark has a power spectrum approximately directly proportional to the host image’s and thereby more robust against Wiener filtering attack.

In this paper, we propose a method to estimate user attention to displayed content signals with temporal analysis of their exhibited behavior. Detecting user attention and controlling contents are key issues in our “networked interaction therapy system” that effectively attracts the attention of memory-impaired people. In our proposed method, user behavior, including body motions (beat actions), is detected with auditory/vision-based methods. This design is based on our observations of the behavior of memory-impaired people under video watching conditions. User attention to the displayed content is then estimated based on body motions synchronized to auditorial signals. Estimated attention levels can be used for content control to attract deeper attention of viewers to the display system. Experimental results suggest that the proposed method effectively extracts user attention to musical signals.

This work explores full registration of scenes in a large area purely based images for city indexing and visualization. Ground-based images including route panoramas, scene tunnels, panoramic views, and spherical views are acquired in the area and are associated with geospatial information. In this paper, we plan distributed locations and paths in the urban area based on the visibility, image properties, image coverage, and scene importance for image acquisition. The criterion is to use a small number of images to cover as large scenes as possible. LIDAR data are used in this view evaluation and real data are acquired accordingly. The extended images realize a compact and complete visual data archiving, which will enhance the perception of spatial relations of scenes.

We developed a novel system consisting of two modules, statistics-based synthesis and non-photorealistic rendering (NPR), to synthesize caricatures of exaggerated facial features and other particular characteristics, such as beards or nevus. The statistics-based synthesis module can exaggerate shapes and positions of facial features based on non-linear exaggerative rates determined automatically. Instead of comparing only the inter relationship between features of different subjects at the existing methods, our synthesis module applies both inter and intra (i.e. comparisons between facial features of the same subject) relationships to make the synthesized exaggerative shape more contrastive. Subsequently, the NPR module generates a line-drawing sketch of original face, and then the sketch is warped to an exaggerative style with synthesized shape points. The experimental results demonstrate that this system can automatically, and effectively, exaggerate facial features, thereby generating corresponding facial caricatures.

In this paper, we propose a method for pose-invariant facial expression recognition from monocular video sequences. The advantage of our method is that, unlike existing methods, our method uses a very simple model, called the variable-intensity template, for describing different facial expressions, making it possible to prepare a model for each person with very little time and effort. Variable-intensity templates describe how the intensity of multiple points defined in the vicinity of facial parts varies for different facial expressions. By using this model in the framework of a particle filter, our method is capable of estimating facial poses and expressions simultaneously. Experiments demonstrate the effectiveness of our method. A recognition rate of over 90% was achieved for horizontal facial orientations on a range of ±40 degrees from the frontal view.