Catálogo de publicaciones - libros

In this paper, we have presented a fully automatic technique for detection and classification of the six basic facial expressions from nearly frontal face images. Facial expressions are communicated by subtle changes in one or more discrete features such as tightening the lips, raising the eyebrows, opening and closing of eyes or certain combinations of them. These discrete features can be identified through monitoring the changes in muscles movement (Action Units) located near about the regions of mouth, eyes and eyebrows. In this work, we have used eleven feature points that represent and identify the principle muscle actions as well as provide measurements of the discrete features responsible for each of the six basic human emotions. A multi-detector approach of facial feature point localization has been utilized for identifying these points of interests from the contours of facial components such as eyes, eyebrows and mouth. Feature vector composed of eleven features is then obtained by calculating the degree of displacement of these eleven feature points from a non-changeable rigid point. Finally, the obtained feature sets are used for training a K-Nearest Neighbor Classifier so that it can classify facial expressions when given to it in the form of a feature set. The developed Automatic Facial Expression Classifier has been tested on a publicly available facial expression database and on an average 90.76% successful classification rate has been achieved.

Digital 3D models of patients’ organs or tissues are often needed for surgical planning and outcome evaluation, or to select prostheses adapted to patients’ anatomy. Tissue classification is one of the hardest problems in automatic model generation from raw data. The existing solutions do not give reliable estimates of the accuracy of the resulting model. We propose a simple generative model using Gaussian Mixture Models (GMMs) to describe the likelihood functions involved in the computation of posterior probabilities. Multiscale feature descriptors are used to exploit the surrounding context of each element to be classified. Supervised learning is carried out using datasets manually annotated by expert radiologists. 3D models are generated from the binary volumetric models, obtained by labelling cortical bone pixels according to maximal likelihoods.

In this paper we present an overview of our novel line segment based watershed segmentation algorithm. Most of the existing watershed algorithms use the region label image as the main data structure for its ease of use. These type of watershed algorithms have a relatively large memory footprint and need unbounded memory access. For our new watershed algorithm we replaced the traditional region label image with a data structure that stores the regions in linked line segments. Consequently, the new algorithm has a much smaller memory footprint. Using the new data structure also makes it possible to create an efficient algorithm that only needs one scan over the input image and that only needs the last 3 lines and a small part of the data structure in memory.

This paper presents a new content based image retrieval approach based on histogram of pattern orientations, namely pattern orientation histogram (POH) method. POH represents the spatial distribution of five different pattern orientations: vertical, horizontal, diagonal down/left, diagonal down/right and non-orientation. In this method, a given image is first divided into image-blocks and frequency of each type of patterns is determined in each image-block. Then local pattern histograms for each of these image-blocks are computed. As a result, local pattern histograms are obtained for each image. We compared our method with one of the texture descriptors MPEG-7 standard. Experimental results demonstrate that POH leads to better precision and recall rates than edge histogram descriptor (EHD) of MPEG-7 standard for search and retrieval of digital imagery.

We describe a novel eye detection method that is robust to the obstacles such as surrounding illumination, hair, and eye glasses. The obstacles above a face image are constraints to detect eye position. These constraints affect the performance of the face applications such as face recognition, gaze tracking, and video indexing systems. To overcome this problem, the proposed method for eye detection consists of three steps. First, the self quotient images are applied to the face images by rectifying illumination. Then, unnecessary pixels for eye detection are removed by using the symmetry object filter. Next, the eye candidates are extracted by using the gradient descent which is a simple and a fast computing method. Finally, the classifier, which has trained by using AdaBoost algorithm, selects the eyes from all of the eye candidates. The usefulness of the proposed method has been demonstrated in an embedded system with the eye detection performance.

This paper describes a novel discrimination method of lung cancers based on statistical analysis of thoracic computed tomography (CT) scans. Our previous Computer-Aided Diagnosis (CAD) system can detect lung cancers from CT scans, but, at the same time, yields many false positives. In order to reduce the false positives, the method proposed in the present paper uses a relationship between lung cancers, false positives and image information on CT scans. The trend of variation of the relationships is acquired through statistical analysis of a set of CT scans prepared for training. In testing, by use of the trend, the method predicts the appearance of lung cancers and false positives in a CT scan, and improves the accuracy of the previous CAD system by modifying the system’s output based on the prediction. The method is applied to 218 actual thoracic CT scans with 386 actual lung cancers. Receiver operating characteristic (ROC) analysis is used to evaluate the results. The area under the ROC curve (Az) is statistically significantly improved from 0.918 to 0.931.