Computer Vision Applications of Self-Organizing Neural Networks

Computer Vision Applications of Self-Organizing Neural Networks

José García-Rodríguez (University of Alicante, Spain), Juan Manuel García-Chamizo (University of Alicante, Spain), Sergio Orts-Escolano (University of Alicante, Spain), Vicente Morell-Gimenez (University of Alicante, Spain), José Antonio Serra-Pérez (University of Alicante, Spain), Anatassia Angelolopoulou (University of Westminster, UK), Alexandra Psarrou (University of Westminster, UK), Miguel Cazorla (University of Alicante, Spain) and Diego Viejo (University of Alicante, Spain)
DOI: 10.4018/978-1-4666-2672-0.ch008

Abstract

This chapter aims to address the ability of self-organizing neural network models to manage video and image processing in real-time. The Growing Neural Gas networks (GNG) with its attributes of growth, flexibility, rapid adaptation, and excellent quality representation of the input space makes it a suitable model for real time applications. A number of applications are presented, including: image compression, hand and medical image contours representation, surveillance systems, hand gesture recognition systems, and 3D data reconstruction.
Chapter Preview
Top

Introduction

The growing computational power of current computer systems and the reduced costs of image acquisition devices allow a very large audience to have an easier access to image analysis issues.

Self-organizing models (SOM), with their massive parallelism, have shown considerable promise in a wide variety of application areas, not related to vision problems only, and have been particularly useful in solving problems for with traditional techniques have failed or proved inefficient. Accordingly, even hard video and image processing applications like some robotic operation, visual inspection, remote sensing, autonomous vehicle driving, automated surveillance, and many others, have been approached using neural networks.

The investigation in SOM in this context has received a great attention, on one hand, as imaging tasks are computationally intensive and high performances potentially can be reached in real time, and on the other hand, thanks to the versatility of neural approaches. Versatility means that a set of interesting properties are shared by neural systems; apart from the inherent parallelism, they allow a distributed representation of the information, easy learning by examples, high generalization ability, and a certain fault tolerance. There is a vast literature on NN, it deals with both theoretical investigation of their inherent mechanisms, and solutions to real problems. SOM have been and have been used extensively for pattern recognition problems, namely classification, clustering, and feature selection.

Several works have used self-organizing models for the representation and tracking of objects. Fritzke (Fritzke, 1997) proposed a variation of the GNG (Fritzke, 1995) to map non–stationary distributions that (Frezza-Buet, 2008). applies to the representation and tracking of people. In [4] it is suggested the use of self-organized networks for human-machine interaction. In [5], amendments to self-organizing models for the characterization of the movement are proposed.

From the works cited, only (Frezza-Buet, 2008) represents both the local movement and the global movement, however there is no consideration of time constraints, and do not exploits the knowledge of previous frames for segmentation and prediction in subsequent frames. Neither uses the structure of the neural network to solve the problem of correspondence in the analysis of the movement.

The time constraints of the problem suggest the need for high availability systems ca-pable to obtain representations with an acceptable quality in a limited time. Besides, the large amount of data suggests definition of parallel solutions.

In this paper we present applications of a neural architecture based on GNG that is able to adapt the topology of the network of neurons to the shape of the entities that appear in the images and can represent and characterize objects of interest in the scenes with the ability to track the objects through a sequence of images in a robust and simple way.

It has been demonstrated that using architectures based on (GNG), can be assumed the temporary restrictions on problems such as tracking objects or the recognition of gestures, with the ability to processing sequences of images offering an acceptable quality of representation and refined very quickly depending on the time available.

With regard to the processing of image sequences, it is possible to accelerate the track-ing and allow the architecture to work at video frequency. In this proposal, the use of the GNG to the representation of objects in sequences solves the costly problem of matching of features over time, using the positions of neurons in the network. Furthermore, can be ensured that from the map obtained with the first image will only be required to represent a fast re-adaptation to locate and track objects in subsequent images, allowing the system to work at video frequency.

The data stored throughout the sequence in the structure of the neural network about characteristics of the entities represented as position, colour, and others, provide information on deformation, merge, paths followed by these entities and other events that may be analyzed and interpreted, giving the semantic description of the behaviors of these entities.

The remainder of the paper is organized as follows: section 2 introduces the capacities of topology learning and preservation of GNG. Section 3 presents a number of video and image processing applications of GNG followed by our major conclusions.

Complete Chapter List

Search this Book:
Reset