Díbio Leandro Borges

Analysis of mammogram classification using a wavelet transform decomposition

In order to fully achieve automated mammogram analysis one has to tackle two problems: classification of radial, circumscribed, microcalcifications, and normal samples; and classification of benign, malign, and normal ones. How to extract and select the best features from the images for classification is a very difficult task, since all of those classes are basically irregular textures with a wide visual variety inside each class. Besides there is a lack of tested solutions for these problems in the literature. In this paper we propose to construct and evaluate a supervised classifier for these two problems, by transforming the data of the images in a wavelet basis, and then using special sets of the coefficients as the features tailored towards separating each of those classes. We have realized that this is a suitable solution worth further exploration. For the experiments we have used samples of images labeled by physicians. Results shown are very promising, and the paper describes possible lines for future directions.

Visual speech recognition: a solution from feature extraction to words classification

Audio-visual speech recognition has been an active area of research lately. A bit, and yet unsolved part of this problem is the visual only recognition, or lip reading. Considering an image sequence of a person pronouncing a word, a full image analysis solution would have to segment the mouth area, extract relevant features, and use them to be able to classify the word from those visual features. We approach this problem by proposing a segmentation technique for the lips contours together with a set of features based on the extracted contours which is able to perform lip reading with promising results. We have collected visual speech sequences in our lab and show the results for a set of ten words in Brazilian Portuguese, spoken by different speakers in more than 150 samples. The approach can be extended and applied to other spoken languages as well.

Automated mammogram classification using a multiresolution pattern recognition approach

In order to fully achieve automated mammogram analysis one has to tackle two problems: classification of radial, circumscribed microcalcifications, and normal samples; and classification of benign, malignant, and normal samples. How to extract and select the best features from the images for classification is a very difficult task, since all of those classes are basically irregular textures with a wide visual variety inside each class. The authors propose a multiresolution pattern recognition approach for this problem, by transforming the data of the images in a wavelet basis, and then using special sets of the coefficients as the features tailored towards separating each of those classes. For the experiments, we have used samples of images labeled by physicians. Results shown are very promising, and the paper describes possible lines for future directions.

Segmentation of postal envelopes for address block location:an approach based on feature selection in wavelet space

This paper presents a segmentation algorithm basedon feature selection in wavelet space. The aim is toautomatically separate in postal envelopes the regionsrelated to background, stamps, rubber stamps, and theaddress blocks. First, a typical image of a postalenvelope is decomposed using Mallat algorithm and Haarbasis. High frequency channel outputs are analyzed tolocate salient points in order to separate the background.A statistical hypothesis test is taken to decide upon moreconsistent regions in order to clean out some noise left.The selected points are projected back to the originalgray level image, where the evidence from the waveletspace is used to start a growing process to include thepixels more likely to belong to the regions of stamps,rubber stamps, and written area. Experiments are runusing original postal envelopes from the Brazilian PostOffice Agency, and here we report results on 440 imageswith many different layouts and backgrounds.

A visual attention model for tracking regions based on color correlograms

The amount of information contained in a single color shot of a scene is extreme considering the variety of tasks that can be performed relying on visual data only. For the purpose of analyzing scenes dynamically, where objects come and go, one has to work under a structure of a visual attention model which prioritizes what type of visual data the system has to follow. This paper presents a novel visual attention model for region tracking based on color correlograms. First, a reference frame is picked and it is segmented for the most significant color regions present in the shot. Color correlograms are then run on every frame after in order to provide spectral and spatial information, for the visual attention control, about the past and the new objects appearing on the scenes. The visual attention (VA) model then keeps track of old and new color regions appearing on the scene until a different frame is chosen. We have run experiments which show the performance of this proposed VA model.

Detecting and Ranking Saliency for Scene Description

There is a long tradition in Computational Vision research regarding Vision as an information processing task which builds up from low level image features to high level reasoning functions. As far as low level image detectors are concerned there is a plethora of techniques found in the literature, although many of them especially designed for particular applications. For natural scenes, where objects and backgrounds change frequently, finding regions of interest which were triggered by a concentration of non-accidental properties can provide a more informative and stable intermediate mechanism for scene description than just low level features. In this paper we propose such a mechanism to detect and rank salient regions in natural and cluttered images. First, a bank of Gabor filters is applied to the image in a variety of directions. The most prominent directions found are then selected as primitive features. Starting from the selected directions with largest magnitudes a resultant is computed by including directional features in the image neighborhood. The process stops when inclusion of other points in the region makes the resultant direction change significantly from the initial one. This resultant is the axis of symmetry of that salient region. A rank is built showing in order the salient regions found in a scene. We report results on natural images showing a promising line of research for scene description and visual attention.

Salient Features and Hypothesis Testing: evaluating a novel approach for segmentation and address block location

This paper presents a modification with further experiments of a segmentation algorithm based on feature selection in wavelet space of ours [9]. The aim is to automatically separate in postal envelopes the regions related to background, stamps, rubber stamps, and the address blocks. First, a typical image of a postal envelope is decomposed using Mallat algorithm and Haar basis. High frequency channel outputs are analyzed to locate salient points in order to separate the background. A statistical hypothesis test is taken to decide upon more consistent regions in order to clean out some noise left. The selected points are projected back to the original gray level image, where the evidence from the wavelet space is used to start a growing process to include the pixels more likely to belong to the regions of stamps, rubber stamps, and written area. We have modified the growing process controlled by the salient points and the results were greatly improved reaching success rate of over 97%. Experiments are run using original postal envelopes from the Brazilian Post Office Agency, and here we report results on 440 images with many different layouts and backgrounds.

A low-cost parallel K-means VQ algorithm using cluster computing

In this paper we propose a parallel approach for the K-meansVector Quantization (VQ) algorithm used in a two-stageHidden Markov Model (HMM)-based system forrecognizing handwritten numeral strings. With thisparallel algorithm, based on the master/slave paradigm,we overcome two drawbacks of the sequential version: a)the time taken to create the codebook; and b) the amountof memory necessary to work with large trainingdatabases. Distributing the training samples over theslaves local disks reduces the overhead associated withthe communication process. In addition, modelspredicting computation and communication time havebeen developed. These models are useful to predict theoptimal number of slaves taking into account the numberof training samples and codebook size.