Invariant 2D object recognition using eigenvalues of covariance matrices, re-sampling and autocorrelation

摘要

This study presents a novel invariant object recognition method for two-dimensional object. The proposed method employs the eigenvalues of covariance matrix, re-sampling, and autocorrelation transformation to extract unique features from boundary information, and then use minimum Euclidean distance method (MD) and backpropagation neural networks (BPN) for classification. The boundary of the binary digital part is first extracted and represented as the sequence of the smaller eigenvalues of covariance matrix over a given region of support. Then the sequence is re-sampled into a pre-determined number, and transformed using autocorrelation function. The experimental results reveal that the proposed method successfully derives translation, rotation, and scaling invariant features which can be classified easily with MD and BPN.