Recognizing polyhedral objects using neural networks

Two different neural networks: the backpropagation (BPNN) and the generalized regression (GRNN) neural networks were used to solve the polyhedral object recognition problem. Some comparisons between these two NNs with the absence and presence of noise-characterized by the absence of significant segments or the presence of spurious ones-were done. The proposed schemes (using either a BPNN or the GRNN) consist of two phases: model building and object recognition. During the model building process, each characteristic view (CV) of the object is described by a feature vector containing the normalized distances from each CV´s vertex to the CV´s centroid. During the object recognition phase, a set of vectors is used to train the NN. Finally a vector representing one of the object´s CVs is presented to the NN to test its performance as a classifier