REDES NEURAIS APLICADAS AO RECONHECIMENTO DE IMAGENS BI-DIMENSIONAIS / USE OF ARTIFICIAL NEURAL NETWORKS IN THE RECOGNITION OF BI-DIMENSION IMAGES

AUTOR(ES)
DATA DE PUBLICAÇÃO

1996

RESUMO

This work investigates the use of Artificial Neural Networks in the recognition of bi-dimensional images. The work was divided in four main parts: a survey on the importance of Computational Vision and on the benefits of the application of intelligent techniques in the fiels; a survey on the structure of image recognition systems found in the literature; the development of two image recognition systems based on neural networks; and an analysis of the performance of the developed systems. Computational Vision has benefited from the main Computational Intelligence techniques (neural networks, genetic algoritms and fuzzy logic) to implement image recognition systems. In this work, the usage of different Kinds of neural networks in image classification was studied: Back-Propagation, Competitive, RBF and Hierarchical. Besiades that, a survey on the fields of application of Computational Vision was made. The basic structure is typically composed of three modules: a pre-processor, a characteristics extractor and a classifier. In this work, two image recognition systems, called Xvision and SimpleNet, were developed. The XVision system follows the structure described above, while the SimpleNet system performs the classification using the information present in the raw picture. The pre-processing module of the Xvision system executes a series of transforms over the image, extracting its essential characteristics so that an invariant representation of the image can be obtained. This pre- processor is based on a previous work in the fiels of Signal Processing. The characteristcs extractor aims to detect the most relevant information present in the image representation obtained after the previous step. Two kinds of extractors were investigated: one based on statistical tecniques (applyng the Fisher`s discriminant) and another based on intelligent techniques (applyng genetic algorithms). The classification module was implementede through several Kinds of neural networks: Back-Propagation, Competitive, RBF and Hierarchical. The pre-processing of the SimpleNet system simply reduces the image`s dimensions. Since the image`s pixels are used for the classification process, no characteristics extractor module was implemented. In the classification module, Back-Propagation and Competitive neural networks were employed. The Xvision system yielded promising results for two sets of objects: the first one composed of mechanical parts and the second one composed of trivial objects. The samples used during the tests presented different characteristics from those samples used during the training process - not only rotated, translated and scaled, but also with structural differences. The classifier obtained a hit ratio above 83% with both sets. The SimpleNet system also showed a good performance in the differentiation of similar objects (telephone cards and X-rays of lungs), achieving hit ratios of more than 80%. The development of both systems demonstrated the viability of the use of neural networks in the classification of bi- dimensional objects. Due to the interest of applying Vision systems in real-time, the time spent in the recognition process was measured. This allowed the detection of the systems` bottlenecks, making their optimization easier.

ASSUNTO(S)

neural networks reconhecimento de imagens computer vision computational intelligence visao computacional redes neurais inteligencia computacional recognition of images

ACESSO AO ARTIGO

Documentos Relacionados