Catálogo de publicaciones - libros

This paper presents a novel texture descriptor based on the wavelet transform. First, we will consider vertical and horizontal coefficients at the same position as the components of a bivariate random vector. The magnitud and angle of these vectors are computed and its histograms are analyzed. This empirical magnitud histogram is modelled by using a gamma distribution (pdf). As a result, the feature extraction step consists of estimating the gamma parameters using the maxima likelihood estimator and computing the circular histograms of angles. The similarity measurement step is done by means of the well-known Kullback-Leibler divergence. Finally, retrieval experiments are done using the Brodatz texture collection obtaining a good performance of this new texture descriptor. We compare two wavelet transforms, with and without downsampling, and show the advantage of the second one, which is translation invariant, for the construction of our texture descriptor.

A local adaptive restoration technique using a sliding discrete cosine transform (DCT) is presented. A minimum mean-square error estimator in the domain of a sliding DCT for image restoration is derived. The local restoration is performed by pointwise modification of local DCT coefficients. To provide image processing in real time, a fast recursive algorithm for computing the sliding DCT is utilized. The algorithm is based on a recursive relationship between three subsequent local DCT spectra. Computer simulation results using a real image are provided and compared with that of common restoration techniques.

This paper builds upon a previous texture feature selection and classification methodology by extending it with two state-of-the-art fami lies of texture feature extraction methods, namely Manjunath & Ma’s Gabor wavelet filters and Local Binary Pattern operators (LBP), which are integrated with more classical families of texture filters, such as co-occur rence matrices, Laws filters and wavelet transforms. Results with Brodatz compositions and outdoor images are evaluated and discussed, being the basis for a comparative study about the discrimination capabilities of those different families of texture methods, which have been traditionally applied on their own.

The lower the resolution of a given text is, the more difficult it becomes to segment and to recognize it. The resolution of screen-rendered text can be very low. With a typical x-height of 4 to 7 pixels it is much lower as in other low resolution OCR situations. Modern OCR approaches for such very low resolution text use a classification-based segmentation where the underlying classifier plays an important role. This paper presents a multiple classifier system for the classification of single characters. This system is used as a subsystem for the classification-based segmentation within a system to read screen-rendered text. The paper shows that the presented multiple classifier system outperforms the best former single classifier system on single characters by far and it shows the impact of using the multiple classifier system on the word reading performance.

In this paper, we propose an approach to classify the film categories by using low-level features and visual features. The goal of this approach is to classify the films into genres. Our current domain of study is the movie preview. A film preview often emphasizes the theme of a film and hence provides suitable information for classification process. In our approach, we classify films into three broad categories: Action, Dramas, and Thriller films. Four computable video features (average shot length, color variance, motion content and lighting key) and visual effects are combined in our approach to provide the advantage information to demonstrate the movie category. Our approach can also be extended for other potential applications, including browsing, retrieval of videos on the internet, video-on-demand, and video libraries.

In content based image retrieval, relevance feedback has been extensively studied to bridge the gap between low level image features and high level semantic concepts. However, it is still challenged by small sample size problem, since users are usually not so patient to label a large number of training instances. In this paper, two strategies are proposed to tackle this problem: (1) a novel active selection criterion. It takes into consideration both the informative and the representative measures. With this criterion, the diversities of the selected images are increased while their informative powers are kept, thus more information gain can be obtained from the feedback images; and (2) incorporation of unlabeled images within the co-training framework. Unlabeled data partially alleviates the training data scarcity problem, thus can improve the efficiency of SVM active learning. Systematic experimental results verify the superiority of our method over some existing active learning methods.

The Nearest Neighbor (NN) rule is one of the simplest and most effective pattern classification algorithms. In basic NN rule, all the instances in the training set are considered the same to find the NN of an input test pattern. In the proposed approach in this article, a local weight is assigned to each training instance. The weights are then used while calculating the adaptive distance metric to find the NN of a query pattern. To determine the weight of each training pattern, we propose a learning algorithm that attempts to minimize the number of misclassified patterns on the training data. To evaluate the performance of the proposed method, a number of UCI-ML data sets were used. The results show that the proposed method improves the generalization accuracy of the basic NN classifier. It is also shown that the proposed algorithm can be considered as an effective instance reduction technique for the NN classifier.

A prior probability model is adapted to a class of images by identification, or parameter estimation from training data. We propose a new and accurate analytical identification of a generic Markov-Gibbs random field (MGRF) model with multiple pairwise interaction and use it for structural analysis and synthesis of textures.

This paper presents a fast multispectral texture defect detection method based on the underlying three-dimensional spatial probabilistic image model. The model first adaptively learns its parameters on the flawless texture part and subsequently checks for texture defects using the recursive prediction analysis. We provide colour textile defect detection results that indicate the advantages of the proposed method.

This paper explores the use of tree-based data structures in shape analysis. We consider a structure which combines several properties of traditional tree models and obtain an efficiently compressed yet faithful representation of shapes. Constructed in a top-down fashion, the resulting trees are unbalanced but resolution adaptive. While the interior of a shape is represented by just a few nodes, the structure automatically accounts for more details at wiggly parts of a shape’s boundary. Since its construction only involves simple operations, the structure provides an easy access to salient features such as concave cusps or maxima of curvature. Moreover, tree serialization leads to a representation of shapes by means of sequences of salient points. Experiments with a standard shape database reveal that correspondingly trained HMMs allow for robust classification. Finally, using spectral clustering, tree-based models also enable the extraction of larger, semantically meaningful, salient parts of shapes.