Catálogo de publicaciones - libros

The conventional technique for scene reconstruction from stereo image pairs searches for the best single surface fitting identified correspondences between the the two images. Constraints on surface continuity, smoothness, and visibility (occlusions) are incorporated into a ‘cost’ – usually an linear combination of signal similarity criteria, with empirically selected coefficients. An unsatisfactory feature of this approach is that matching accuracy is very sensitive to correct choice of these coefficients. Also, few real scenes have only one surface, so that the single surface assumption contributes to matching errors.

We propose a noise-driven paradigm for stereo matching that does not couple the matching process with choice of surfaces by imposing constraints in the matching step. We call our strategy ‘Concurrent Stereo Matching’ because the first step involves a high degree of parallelism (making real-time implementations possible using configurable hardware): rather than search for ‘best’ matches, it first identifies all 3D volumes that match within a criteria based on noise in the image. Starting in the foreground, these volumes are then examined and surfaces are selected which exhibit high signal similarity in both images. Local constraints on continuity and visibility – rather than global ones – are used to select surfaces from the candidates identified in the first step.

Object recognition under occlusions is an important problem in computer vision, not yet completely solved. In this note we describe a simple but effective technique for the recognition objects under occlusions. The proposal uses the most distinctive parts of the objects for their further detection. During training, the proposal, first detects the distinctive parts of each object. For each of these parts an invariant description in terms of invariants features is next computed. With these invariant descriptions a specially designed set of associative memories (AMs) is trained. During object detection, the proposal, first looks for the important parts of the objects by means of the already trained AM. The proposal is tested with a bank of images of real objects and compared with other similar reported techniques.

In this paper, a fully automatic, real-time system is proposed to recognize seven basic facial expressions (angry, disgust, fear, happiness, neutral, sadness and surprise). First, faces are located and normalized based on an illumination insensitive skin model and face segmentation; then, the Local Binary Patterns (LBP) techniques, which are invariant to monotonic grey level changes, are used for facial feature extraction; finally, the Linear Programming (LP) technique is employed to classify seven facial expressions. Theoretical analysis and experimental results show that the proposed system performs well in some degree of illumination changes and head rotations.

In this document we present an agent for people detection and tracking through stereo vision. The agent makes use of the active vision to perform the people tracking with a robotic head on which the vision system is installed. Initially, a map of the surrounding environment is created including its motionless characteristics. This map will later on be used to detect objects in motion, and to search people among them by using a face detector. Once a person has been spotted, the agent is capable of tracking them through the robotic head that allows the stereo system to rotate. In order to achieve a robust tracking we have used the Kalman filter. The agent focuses on the person at all times by framing their head and arms on the image. This task could be used by other agents that might need to analyze gestures and expressions of potential application users in order to facilitate the human-robot interaction.

This paper shows a methodology for on-line recognition and classification of pieces in robotic assembly tasks and its application into an intelligent manufacturing cell. The performance of industrial robots working in unstructured environments can be improved using visual perception and learning techniques. The object recognition is accomplished using an Artificial Neural Network (ANN) architecture which receives a descriptive vector called CFD&POSE as the input. This vector represents an innovative methodology for classification and identification of pieces in robotic tasks, every stage of the methodology is described and the proposed algorithms explained. The vector compresses 3D object data from assembly parts and it is invariant to scale, rotation and orientation, and it also supports a wide range of illumination levels. The approach in combination with the fast learning capability of ART networks indicates the suitability for industrial robot applications as it is demonstrated through experimental results.

This paper presents a technique for color image segmentation, product of the combination and improvement of a number of traditional approaches: Seed region growth, Threshold classification and level on detail in the analysis of demand. First, a set of precise color classes with variable threshold is defined based on sample data. A scanline algorithn uses color clases with a small threshold to extract an initial group of pixels. These pixels are passed to a region growth method, which performs segmentation using higher-threshold classes as homogeneity criterion to stop growth. This hybrid technique solves disadvantages from individual methods and keeps their strengths. Its advantages include a higher robustness to external noise and variable illumination, efficiency on image processing, and quality on region segmentation, outperforming the results of standalone implementations of individual techniques. In addition, the proposed approach sets a starting point for further improvements.

The recognition of images beyond basic image processing often relies on training an adaptive system using a set of samples from a desired type of images. The adaptive algorithm used in this research is a learning automata model called CLS (collective learning systems). Using CLS, we propose a hierarchy of collective learning layers to learn color and texture feature patterns of images to perform three basic tasks: recognition, classification and segmentation. The higher levels in the hierarchy perform recognition, while the lower levels perform image segmentation. At the various levels the hierarchy is able to classify images according to learned patterns. In order to test the approach we use three examples of images: a) Satellite images of celestial planets, b) FFT spectral images of audio signals and c) family pictures for human skin recognition. By studying the multi-dimensional histogram of the selected images at each level we are able to determine the appropriate set of color and texture features to be used as input to a hierarchy of adaptive CLS to perform recognition and segmentation. Using the system in the proposed hierarchical manner, we obtained promising results that compare favorably with other AI approaches such as Neural Networks or Genetic Algorithms.

(1909-1997)

In this paper we present the performance evaluation of a texture segmentation approach for synthetic textured images. Our segmentation approach uses a Bayesian inference procedure using co-ocurrence properties over a set of randomly sampled points in the image. We developed an exhaustive performance test for this approach that compares segmentation results to the “ground truth” images under a varying number of sampled points, in the neighborhood of each pixel used to classify it in the test images. We show our preliminary results that let us to choose the optimal number of points to analyze in the neighborhood of each pixel to assign a texture label. This method can be easily applied to segment outdoor real textured images.

A new image retrieval method based on salient points extracted from DCT compressed domain is proposed in this paper. Using significant DCT coefficients, we provide a robust self-adaptive salient point extraction algorithm which is very robust to most of common image processing. Based on salient points, two local image features, color histogram and LBP histogram are computed to represent local properties of the image for retrieval. Our system reduces the amount of data to be processed and only needs to do partial decompression, so it can accelerate the work of image retrieval. The experimental results also demonstrate it improves performance both in retrieval efficiency and effectiveness.

The ISOMAP algorithm has recently emerged as a promising dimensionality reduction technique to reconstruct nonlinear low-dimensional manifolds from the data embedded in high-dimensional spaces, by which the high-dimensional data can be visualized nicely. One of its advantages is that only one parameter is required, i.e. the neighborhood size or K in the K nearest neighbors method, on which the success of the ISOMAP algorithm depends. However, it’s an open problem how to select a suitable neighborhood size. In this paper, we present an effective method to select a suitable neighborhood size, which is much less time-consuming than the straightforward method with the residual variance, while yielding the same results. In addition, based on the characteristics of the Euclidean distance metric, a faster Dijkstra-like shortest path algorithm is used in our method. Finally, our method can be verified by experimental results very well.