Título: Morphological Identification Of Inductive Signatures Using An Adaptive Resonance Theory Based Learning Scheme
Autores: Lima, Glauston R. Teixeira de; Silva, José Demísio S.; Saotome, Osamu
Resumo: This paper presents an algorithm that uses a learning scheme based on the Adaptive Resonance Theory – ART2 to identify morphological patterns in two classes of vehicle inductive signatures and to construct inductive signatures subclasses based on morphological similarities among inductive signatures and previously identified patterns. The algorithm is initialized using all inductive signatures of a class as inputs to the learning scheme. During the iterations of the algorithm, the inputs whose morphological similarity exceeds a pre-established threshold value are merged into a single new input to the learning scheme. The merging process stops when the threshold value can not be overcome. The remaining inputs are considered to be the desired group of patterns representing the morphological diversity of the inductive signature class. Next, the morphological patterns identified by the learning scheme are submitted to a clustering algorithm to form the subclasses. In some classification tasks, the problem of classes overlap in the input space can be approached under the point of view of the morphological similarity among groups of patterns of each class. In these cases, pre-processing the original input space dividing it into morphological subclasses serves to identify the regions where the classes overlap is concentrated and to break down the initial classification task, whose solution is more difficult, in minor tasks, hypothetically, easier to be resolved. However, the supposition that the division in subclasses led to a better separation of the two classes of inductive signatures will be checked in subsequent stages of our research through tests with neural classifiers.
Palavras-chave: Vehicle Inductive Signatures; Clustering; Adaptive Resonance Theory Neural Network
Páginas: 16
Código DOI: 10.21528/lmln-vol5-no1-art3
Artigo em PDF: vol5-no1-art3.pdf
Arquivo BibTex: vol5-no1-art3.bib