Catálogo de publicaciones - libros

This article presents a method aiming at filtering objectionable image contents. This kind of problem is very similar to object recognition and image classification. In this paper, we propose to use Adaptive-Subspace Self-Organizing Maps (ASSOM) to generate invariant pornographic features. To reach this goal, we construct local signatures associated to salient patches according to adult and benign databases. Then, we feed these vectors into each specialized ASSOM neural network. At the end of the learning step, each neural unit is tuned to a particular local signature prototype. Thus, each input image generates two neural maps that can be represented by two activation vectors. A supervised learning is finally done by a Normalized Radial Basis Function (NRBF) network to decide the image category. This scheme offers very promising results for image classification with a percentage of 87.8% of correct classification rates.

Aspect-based relevance learning is a relevance feedback scheme based on a natural model of relevance in terms of image aspects. In this paper we propose a number of active learning and interaction strategies, capitalizing on the transparency of the aspect-based framework. Additionally, we demonstrate that, relative to other schemes, aspect-based relevance learning upholds its retrieval performance well under feedback consisting mainly of example images that are only partially relevant.

The problem of face detection remains challenging because faces are non-rigid objects that have a high degree of variability with respect to head rotation, illumination, facial expression, occlusion, and aging. A novel technique that is gaining in popularity, known as Locally Linear Embedding (LLE), performs dimensionality reduction on data for learning and classi-fication purposes. This paper presents a novel approach to the face detection problem by applying the LLE algorithm to 2D facial images to obtain their representation in a sub-space under the specific conditions stated above. The low-dimensional data are then used to train Support Vector Machine (SVM) classifiers to label windows in images as being either face or non-face. Six different databases of cropped facial images, corresponding to variations in head rotation, illumination, facial expression, occlusion and aging, were used to train and test the classifiers. Experimental results obtained demonstrated that the performance of the proposed method was similar and sometimes better when compared to other face detection methods, while using a lower amount of training images, thus indicating a viable and accurate technique.

In this paper, we present a ROI (Region-Of-Interest)-based medical image retrieval system that is considering combination of feature descriptors and initial weights for similarity matching. For semantic ROI segmentation, we create attention window (AW) to remove the meaningless regions included in the image such as background and propose a quad-tree based ROI segmentation method. In addition, in order to improve the retrieval performance and consider human perception, initial weights for feature distances are also proposed. From, several experiments, we demonstrate that the ROI-based method having different initial weights shows the better performance than previous related methods.

Rushes footage are considered as cheap gold mine with the potential for reuse in broadcasting and filmmaking industries. However, it is difficult to mine the “gold” from the rushes since usually only minimum metadata is available. This paper focuses on the structuring and indexing of the rushes to facilitate mining and retrieval of “gold”. We present a new approach for rushes structuring and indexing based on motion feature. We model the problem by a two-level Hierarchical Hidden Markov Model (HHMM). The HHMM, on one hand, represents the semantic concepts in its higher level to provide simultaneous structuring and indexing, on the other hand, models the motion feature distributions in its lower level to support the encoding of the semantic concepts. The encouraging experimental results on TRECVID′05 BBC rushes demonstrate the effectiveness of our approach.

The use of local features in computer vision has shown to be promising. Local features have several advantages including invariance to image transformations, independence of the background, and robustness in difficult situations like partial occlusions. In this paper we suggest using local integral invariants to extract local image descriptors around interest points and use them for the retrieval task. Integral invariants capture the local structure of the neighborhood around the points where they are computed. This makes them very well suited for constructing highly-discriminative local descriptors. We study two types of kernels used for extracting the feature vectors and compare the performance of both. The dimensionality of the feature vector to be used is investigated. We also compare our results with the SIFT features. Excellent results are obtained using a dataset that contains instances of objects that are viewed in difficult situations that include clutter and occlusion.

Efficient image retrieval from large image databases is a challenging problem. In this paper we present a method offering constant time complexity for the comparison of two shapes. In order to achieve this, we extend the qualitative concept of positional-contrast by 86 new relations describing the position of a polygon w. r. t. its line segments. On this basis a histogram of the relations’ frequencies is computed for each shape. A useful property of our approach is that, due to the underlying concept of positional-contrast, it can be intuitively decided whether its combination with other features is promising. Especially, retrieval results of about 64% are achieved in the MPEG test with constant time complexity.

This paper relates learning rare concepts for multimedia retrieval to a more general setting of imbalanced data. A Relay Boost (RL.Boost) algorithm is proposed to solve this imbalanced data problem by fusing multiple features extracted from the multimedia data. As a modified RankBoost algorithm, RL.Boost directly minimizes the ranking loss, rather than the classification error. RL.Boost also iteratively samples positive/negative pairs for a more balanced data set to get diverse weak ranking with different features, and combines them in a ranking ensemble. Experiments on the standard TRECVID 2005 benchmark data set show the effectiveness of the proposed algorithm.

Most of the researchers have used spatio-temporal relations for retrieval in video. It’s just trajectory-based or content-based retrieval. However, we seldom retrieve information referring to semantics. So, in this paper, we propose a novel approach for motion recognition from the aspect of semantic meaning. This issue can be addressed through a hierarchical model that explains how the human language interacts with human motions. And, in the experiment part, we evaluate our new approach using trajectory distance based on spatio-temporal relations to distinguish the conceptual similarity and get the satisfactory results.

Traditional image-based face detection methods use a window based scanning technique where the window is scanned pixel-by-pixel to search for faces in various positions and scales within an image. Therefore, they require high computation cost and are not adequate to the real time applications. In this paper, we introduce a novel coarse-to-fine classification method for image-based face detection using multiple face classifiers. A coarse location of a face is first classified by the gradient feature based face classifier where the window is scanned in large moving steps. From the coarse location of a face, the fine classification is performed to identify the local image as a face using the multiple face classifiers where the window is finely scanned. The multiple face classifiers are designed to take gradient, texture and pixel intensity features and trained by back propagation learning algorithm. Experimental results demonstrate that our proposed method can reduce up to 90.4% of the number of scans compared to the exhaustive full scanning technique and provides the high detection rate.