Catálogo de publicaciones - libros

We are at the beginning of the digital Age of Information, a digital Renaissance allowing us to communicate, share, and learn in novel ways and resulting in the creation of new paradigms. However, having access to all of the knowledge in the world is pointless without a means to search for it. Visual information retrieval is poised to give access to the myriad forms of images and video, comprising knowledge from individuals and cultures to scientific fields and artistic communities. In this paper I summarize the current state of the field and discuss promising future directions and grand challenges.

The paper presents a framework for information retrieval in visual databases containing colour images. The concept of is adapted to colour images; building and detection of such keypoints are explained in details. The issues of matching images are only briefly highlighted. Finally, the idea of higher-level keypoints is proposed.

An approach for visual information analysis and classification is presented. It is based on a knowledge synthesizing technique to automatically create a relevance map from essential areas in natural images. It also derives a set of well-structured representations from low-level description to drive the final classification. The backbone of this approach is a distribution mapping strategy involving a knowledge synthesizing module based on an intelligent growing when required network. Classification is achieved by simulating the high-level top-down visual information perception in primates followed by incremental Bayesian parameter estimation. The proposed modular system architecture offers straightforward expansion to include user relevance feedback, contextual input, and multimodal information if available.

Image similarity is an important topic in the field of image processing. The goal is to obtain objective measures that express the similarity between two images in a way that matches human evaluation. Such measures have both theoretical and practical applications. In this paper, we show how similarity measures for fuzzy sets have been modified in order to be applied in image processing. We also discuss a new application of these measures in the context of color image retrieval, indicating the potential of this class of similarity measures.

When accessing contents in ever-increasing multimedia chunks, indexing and analysis of video data are key steps. Among different types of videos, sports video is an important type of video and it is under research focus now. Due to the increasing demands from audience, highlights extraction become meaningful. This paper proposed a mean shift clustering based semi-automatic sports video highlight annotation method. Specifically, given small pieces of annotated highlights, by adopting Mean Shift clustering and earth mover’s distance (EMD), mid-level features of highlight shots are extracted and utilized to annotate other highlights automatically. There are 3 steps in the proposed method: First, extract signature of different features – Camera Motion Signature (CMS) for motion and Pivot Frame Signature (PFS) for color. Second, Camera motion’s co-occurrence value is defined as Camera Motion Devotion Value (CMDV) and calculated as EMD distance between signatures. Decisive motion feature for highlights’ occurrences is thus semi-automatically detected. Finally highlights are annotated based on these motion parameters and refined by color-based results. Another innovation of this paper is to combine semantic information with low-level feature aiding highlight annotation. Based on Highlight shot feature (HSF), we performed hierarchical highlight annotation and got promising results. Our method is tested on four video sequences comprising of different types of sports games including diving, swimming, and basketball, over 50,000 frames and experimental results demonstrate the effectiveness of our method.

Face image can be seen as a complex visual object, which combines a set of characterizing facial features. These facial features are crucial hints for machine to distinguish different face images. However, the face image also contains certain amount of redundant information which can not contribute to the face image retrieval task. Therefore, in this paper we propose a retrieval system which is aim to eliminate such effect at three different levels. The Ternary Feature Vector (TFV) is generated from quantized block transform coefficients. Histograms based on TFV are formed from certain subimages. Through this way, irrelevant information is gradually removed, and the structural and statistical information are combined. We testified our ideas over the public face database FERET with the Cumulative Match Score evaluation. We show that proper selection of subimage and feature vectors can significantly improve the performance with minimized complexity. Despite of the simplicity, the proposed measures provide results which are on par with best results using other methods.

In this study, we propose a simple, yet general and powerful framework for constructing accurate affine invariant regions and use it for near-duplicate detection problem. In our framework, a method for extracting reliable seed points is first proposed. Then, regions which are invariant to most common affine transformations are extracted from seed points by a new method named the Thresholding Seeded Growing Region (TSGR). After that, an improved ellipse fitting method based on the Direct Least Square Fitting (DLSF) is used to fit the irregularly-shaped contours of TSGRs to obtain ellipse regions as the final invariant regions. At last, SIFT-PCA descriptors are computed on the obtained regions. In the experiment, our framework is evaluated by retrieving near-duplicate in an image database containing 1000 images. It gives a satisfying result of 96.8% precision at 100% recall.

Focused places of a photo act as a significant cue for image concept discovery and quality assessment. Therefore, to find them is an important issue. In this paper, we design a focusing degree detector by which a focusing degree map is generated for a photograph. The results could be used to obtain focused places of photographs. As a concrete example of their applications, image retrieval and image quality assessment are investigated in this work. The experimental results show that the retrieval algorithm based on this detector and map can get more accurate retrieval results and the proposed assessment algorithm has a high ability to discriminate photos from low quality to high quality.

Content based image retrieval (CBIR) has been researched for decades. However, the "semantic gap" which exists between low-level features and human semantics still remains an unsolved problem. Region based image retrieval (RBIR) was proposed to bridge this gap in some extent. Beyond the pixel values in the image, what other information can also be used? The other information we use is Exif, which records the snapping condition of camera metadata. In this paper we propose an method that combines region low level features of image and camera metadata for image retrieval. Experimental results show the efficiency of our method than the traditional CBIR.

Automatic image annotation aims at labeling images with keywords. In this paper we investigate three annotation benchmark tasks used in literature to evaluate annotation systems’ performance. We empirically compare the first two of the tasks, the 5000 Corel images and the Corel categories tasks, by applying a family of annotation system configurations derived from our PicSOM image content analysis framework. We establish an empirical correspondence of performance levels in the tasks by studying the performance of our system configurations, along with figures presented in literature. We also consider ImageCLEF 2006 Object Annotation Task that has earlier been found difficult. By experimenting with the data, we gain insight into the reasons that make the ImageCLEF task difficult. In the course of our experiments, we demonstrate that in these three tasks the PicSOM system—based on fusion of numerous global image features—outperforms the other considered annotation methods.