Título: An Experimentation with Improved Target Vectors for MLP in Classifying Degraded Patterns
Autores: Nomura, Shigueo; Manzan, José Ricardo Gonçalves; Yamanaka, Keiji
Resumo: In this paper, we adopt unconventional target vectors to improve the performance of pattern classification systems using neural network techniques based on MLP. Instead of conventional target vectors, the new target vectors are bipolar, or- thogonal, and highly dimensional. Since they are orthogonal with bipolar representation, we can take advantage of increasing on the Euclidean distance for these vectors when their number (n in a Euclidean space Rn ) of components increases. We define non-orthogonal bipolar vectors considered as conventional target vectors for comparison purposes. Those non-orthogonal bipolar vectors provide a fair reference to ensure the effectiveness of the adopted unconventional target vectors and in justifying the cred- ibility and validity of experimental results. The conventional and unconventional target vectors are used in the experiments for training MLP models by backpropagation algorithm to classify patterns extracted from actual degraded images. Comparison of experimental results lead to conclusions that classification performances of MLP models considerably improved with the adopted unconventional target vectors in classifying degraded patterns.
Palavras-chave: Multilayer perceptron model; conventional target vector; orthogonal bipolar vector; non-orthogonal bipolar vector; degraded image; pattern recognition
Páginas: 13
Código DOI: 10.21528/lmln-vol8-no4-art5
Artigo em PDF: vol8-no4-art5.pdf
Arquivo BibTex: vol8-no4-art5.bib