Abstract

Pattern clustering algorithms or similarity measures have been used to classify extracted pattern features among reference classes. A new approach is to use properly trained linear associative memory mapping (LAMM). Associative recall may in general be defined as a mapping in which a finite number of input vectors is transformed into a given set of output vectors. In the case of erroneous input vectors, it has been shown that this mapping is least-squares sense optimal. This error tolerance suggests its applicability to pattern classification. For the purpose of classification, the inputs are feature vectors derived using some algorithm, where the outputs are codes (tags) for each pattern. The simplest code vectors could be orthonormal unity vectors. The associative memory matrix is derived during a training process using reference feature vectors and their corresponding code vectors. For texture classification, features can be extracted using a proposed algorithm based on normalized Hough transform. These features remain invariant under geometrical transformation such as rotation, translation, or scaling.

© 1987 Optical Society of America