Narrow Results

Localizing facial features is a critical component in many computer vision applications such as expression recognition, face recognition, face tracking, animation, and red-eye correction. Practical applications require detectors that operate reliably under a wide range of conditions, including variations in illumination, pose, ethnicity, gender and age. One challenge for the development of such detectors is the inherent trade-off between robustness and precision. Robust detectors tend to provide poor localization and detectors sensitive to small changes in local structure, which are needed for precise localization, generate a large number of false alarms. Here we present an approach to this trade-off based on context dependent inference. First, robust detectors are used to detect contexts in which target features occur, then precise detectors are trained to localize the features given the detected context. This paper describes the approach and presents a thorough empirical examination of the parameters needed to achieve practical levels of performance, including the size of the training database, size of the detector's receptive fields and methods for information integration. The approach operates in real time and achieves, to our knowledge, the most accurate localization performance to date.

The bags of feature and feature vocabulary based approaches have been presented for image categorization due to their simplicity and competitive performance. Some modified versions have been subsequently proposed, incorporating the methods such as adapted vocabularies, fast indexing, and Gaussian mixture models. In this paper, we propose an improvement of replacing the Harris-affine detection method by a random sampling procedure together with an increased number of sample points. Experimental results show that this new method improves categorization accuracy on a five-category problem using the Caltech-4 dataset. It is concluded that random sampling produces higher attainable point density and better categorization performance.

Backpropagation is one of the most widely used methods for training multilayer neural networks, yet questions still exist regarding how the networks organize internally during (raining to represent the external training environment. This paper presents empirical measurements showing that feedforward networks, when trained on many separable and non-separable problems, learn a characteristic internal representation, herein called the Network Linear Transform (NLT), that is independent of: (a) the initial weights and cell biases, and (b) the number of hidden units. The internal decomposition (defined as the values of cell weights and biases) of the trained nets, however, is greatly dependent upon these quantities. For the case of orthogonal input patterns, the NLT captures a literal image of the training environment, while for linearly-independent and separable linearly-dependent training sets, the NLT: (a) captures characteristic features of the correct input patterns, (b) captures inverted versions of characteristic features of incorrect patterns, and (c) rejects features common to all pattern classes. For non-separable problems, the NLT captures statistical ensemble information about patterns in each training class. The hidden units act as difference detectors and thus convey information that distinguishes input patterns from one another. They separate the patterns into groups that are easily discriminated by the output cells. A linearized mathematical network model is developed that accurately reproduces weight matrices and cell responses for certain separable learning situations, and which supports the experimental findings given above.

In Part I, we described an efficient and robust algorithm for computing a common refinement of two surface meshes. In this paper, we present a theoretical verification of the robustness of our algorithm by showing the topological preservation of the intersection principle, which we used to resolve topological inconsistencies caused by numerical errors. To enhance robustness in practice for complex geometries, we further propose techniques to detect and match geometric features, such as ridges, corners, and nonmatching boundaries. We report experimental results using our enhanced overlay algorithm with feature matching for complex geometries from real-world applications.

Electrocardiography deals with the electrical activity of the heart. The condition of cardiac health is given by the electrocardiogram (ECG). ECG analysis is one of the most important aspects of research in the field of biomedical sciences and healthcare. The precision in the identification of various parameters in the ECG is of great importance for the reliability of an automated ECG analyzing system and diagnosis of cardiac diseases. Many algorithms have been developed in the last few years, each with their own advantages and limitations. In this work, we have developed an algorithm for 12-lead ECG parameter detection which works in three steps. Initially, the signal is denoised by the wavelet transform approach using a graphical programming language called LabVIEW (Laboratory Virtual Instrument Engineering Workbench). Next, primary features are detected from the denoised ECG signal using Matlab, and lastly, the secondary features related to diabetes are estimated from the detected primary features. Diabetes mellitus (DM), which is characterized by raised blood glucose levels in an individual, affects an estimated 2–4% of the world's population, making it one of the major chronic illnesses prevailing today. Recently, there has been increasing interest in the study of relationship between diabetes and cardiac health. Thus, in this work, we estimate diabetic-related secondary ECG features like corrected QT interval (QTc), QT dispersion (QTd), P wave dispersion (PD), and ST depression (STd). Our software performance is evaluated using CSE DS-3 multi-lead data base and the data acquired at SGGS IE & T, Nanded, MS, which contains 5000 samples recorded at a sampling frequency of 500 HZ. The proposed algorithm gives a sensitivity of 99.75% and a specificity of 99.83%.

In [Westhoff et al., 2005], we proposed a novel method to determine illumination-invariant features in images. The quantitative bilateral symmetry of a given scene is computed using dynamic programming before applying the resulting symmetry image and feature vector signatures to an omnidirectional motion tracking task. Compared to other approaches to symmetry detection in computer vision, the results of our novel algorithm describe symmetry in terms of an absolute region instead of a relative degree. In this paper, we describe further optimizations and details of the proposed Dynamic Programming Symmetry feature detection algorithm (DPS algorithm).

This paper addresses the problem of highly accurate, highly speedy, more reliable and fully automatic camera calibration. Our objective is to construct a reliable and fully automatic system to supply a more robust and highly accurate calibration scheme. A checkerboard pattern is used as calibration pattern. After the corner points on image are detected, an improved Delaunay triangulation based algorithm is used to make correspondences between corner points on image and corner points on checkerboard in 3D space. In order to determine precise position of the actual corner points, a geometrical constraint based global curve fitting algorithm has been developed. The experimental results show that the geometrical constraint based method can improve remarkably the performance of the feature detection and camera calibration.

Two-dimensional Electrophoresis Gel Images Scan (2dEGIS) implements a mix of computational methods for processing two-dimensional electrophoresis gel images. For advancing the analysis in case-control sample studies, a multi-component decomposition-approximation approach is presented, based on: (1) A global scan aimed to detect discriminative patterns with just a few components; (2) A more localized image scan through aggregated components; (3) The exploration of specific regions with maximal localization power. The tool 2dEGIS represents a novel unifying instrument for the computational analysis of gel images.

Backpropagation is one of the most widely used methods for training multilayer neural networks, yet questions still exist regarding how the networks organize internally during training to represent the external training environment. This paper presents empirical measurements showing that feedforward networks, when trained on many separable and non-separable problems, learn a characteristic internal representation, herein called the Network Linear Transform (NLT), that is independent of: (a) the initial weights and cell biases, and (b) the number of hidden units. The internal decomposition (defined as the values of cell weights and biases) of the trained nets, however, is greatly dependent upon these quantities. For the case of orthogonal input patterns, the NLT captures a literal image of the training environment, while for linearly-independent and separable linearly-dependent training sets, the NLT: (a) captures characteristic features of the correct input patterns, (b) captures inverted versions of characteristic features of incorrect patterns, and (c) rejects features common to all pattern classes. For non-separable problems, the NLT captures statistical ensemble information about patterns in each training class. The hidden units act as difference detectors and thus convey information that distinguishes input patterns from one another. They separate the patterns into groups that are easily discriminated by the output cells. A linearized mathematical network model is developed that accurately reproduces weight matrices and cell responses for certain separable learning situations, and which supports the experimental findings given above.