Hugo Vieira Neto

Visual novelty detection with automatic scale selection

This paper presents experiments with an autonomous inspection robot, whose task was to highlight novel features in its environment from camera images. The experiments used two different attention mechanisms-saliency map and multi-scale Harris detector-and two different novelty detection mechanisms - the Grow-When-Required (GWR) neural network and an incremental Principal Component Analysis (PCA). For all mechanisms we compared fixed-scale image encoding with automatically scaled image patches. Results show that automatic scale selection provides a more efficient representation of the visual input space, but that performance is generally better using a fixed-scale image encoding.

Fingerprint classification with neural networks

This paper presents some intermediate results on fingerprint classification adopting a neural network as decision stage, in order to evaluate the performance of a discrete wavelet transform as feature extraction technique. Some issues on the image acquisition, preprocessing and segmentation are also discussed.

Real-time Automated Visual Inspection using Mobile Robots

We present a framework to perform novelty detection using visual input in which a mobile robot first learns a model of normality in its operating environment and later uses this to highlight uncommon visual features that may appear. This ability is of great importance for both robotic exploration and inspection tasks, because it enables the robot to allocate computational and attentional resources efficiently to those features which are novel. At the heart of the proposed system is the image encoding mechanism which uses local colour statistics from regions selected by a biologically-inspired model of visual attention. Our approach works in real-time with a wide, unrestricted field of view and is robust to image transformations. Experiments conducted in an engineered scenario demonstrate the efficiency and functionality of our method.

Automated Exploration and Inspection: Comparing Two Visual Novelty Detectors

Mobile robot applications that involve exploration and inspection of dynamic environments benefit, and often even are dependant on reliable novelty detection algorithms. In this paper we compare and discuss the performance and functionality of two different on-line novelty detection algorithms, one based on incremental Principal Component Analysis and the other on a Grow-When-Required artificial neural network. A series of experiments using visual input obtained by a mobile robot interacting with laboratory and real-world environments demonstrate and measure advantages and disadvantages of each approach.

Biometric-oriented Iris Identification Based on Mathematical Morphology

A new method for biometric identification of human irises is proposed in this paper. The method is based on morphological image processing for the identification of unique skeletons of iris structures, which are then used for feature extraction. In this approach, local iris features are represented by the most stable nodes, branches and end-points extracted from the identified skeletons. Assessment of the proposed method was done using subsets of images from the University of Bath Iris Image Database (1000 images) and the CASIA Iris Image Database (500 images). Compelling experimental results demonstrate the viability of using the proposed morphological approach for iris recognition when compared to a state-of-the-art algorithm that uses a global feature extraction approach.

Classification of Events in Distribution Networks using Autonomous Neural Models

This paper presents a method for automatic clas- sification of faults and events related to quality of service in electricity distribution networks. The method consists in preprocessing event oscillographies using the wavelet transform and then classifying them using autonomous neural models. In the preprocessing stage, the energy present in each sub-band of the wavelet domain is computed in order to compose input feature vectors for the classification stage. The classifiers investigated are based in Multi-Layer Perceptron (MLP) feed-forward artificial neural networks and Support Vector Machines (SVM), which automatically promote input selection and structure complexity control simultaneously. Experiments using simulated data show promising results for the proposed application.

An Accurate Approach for Automatic Segmentation of Power Distribution Voltage Waveforms

This paper addresses one of the fundamental steps in automatic waveform analysis: transient segmentation. We present a new approach which incorporates the advantages of a multilevel wavelet decomposition and the representation of the support vector data description. Real data from a monitoring system developed for lightning overvoltage detection in overhead distribution power lines was used for comparison and validation of segmentation performance. The experiments involve the proposed segmentation approach and usual segmentation methods, such as Kalman filtering, autoregressive models, and standard discrete wavelet transform. The results show that the proposed segmentation method based on DWT+SVDD yields better overall accuracy for transient segmentation when compared to currently used methods, demonstrating the potential for applications in oscillographic recorders for smart distribution networks, where identification, characterization, and mitigation of events are critical for network operation and maintenance.

A new approach for event classification and novelty detection in power distribution networks

This paper presents a new approach for automatic oscillography classification in distribution networks, including the detection of patterns not initially presented to the classifier during training, which are defined as novelties. We performed experiments with coupled novelty detection and multi-class classification, and also in separate stages, using the following classifiers: Gaussian Mixture Models (GMM), K-means clustering (KM), K-nearest neighbors (KNN), Parzen Windows (PW), Support Vector Data Description (SVDD), and multi-class classification based on Support Vector Machines (SVM). Preliminary results for simulated data in the Alternative Transient Program (ATP) demonstrate the ability of the method to identify new classes of events in a dynamic learning environment. This work was partially supported by COPEL within the Research and Development Program of the Brazilian Electrical Energy Agency (ANEEL).

A Study of the Effect of Illumination Conditions and Color Spaces on Skin Segmentation

This work aims at investigating the influence of luminance information and environment illumination on skin classification. We explore Bayesian approaches to perform automatic classification of human skin pixels on digital images, using color features as input. Two probabilistic skin color models were built on different color spaces (RGB, normalized RG, HSI, HS, YCbCr and CbCr) and tested in a task of automatic pixel classification into skin and non-skin. Analyses of classification performance were done by presenting an illumination controlled image database containing images acquired in four different illumination conditions (shadow, sun, incandescent and fluorescent lights) to these classifiers. Our experiments show that building probabilistic skin color models using the CbCr color space generally improves performance of the classifiers and that best performance is achieved in shadow illumination.

Fast Saliency Detection Using Sparse Random Color Samples and Joint Upsampling

The human visual system employs a mechanism of visual attention, which selects only part of the incoming information for further processing. Through this mechanism, the brain avoids overloading its limited cognitive capacities. In computer vision, this task is usually accomplished through saliency detection, which outputs the regions of an image that are distinctive with respect to its surroundings. This ability is desirable in many technological applications, such as image compression, video quality assessment and content-based image retrieval. In this paper, a saliency detection method based on color distance with sparse random samples and joint upsampling is presented. This approach computes full-resolution saliency maps with short runtime by leveraging both edge-preserving smoothing and joint upsampling capabilities of the Fast Global Smoother. The proposed method is assessed through precision-recall curves, F-measure and average runtime on the MSRA1K dataset. Results show that the method is competitive with state-of-the-art algorithms in both saliency detection accuracy and runtime.