STRICR-FB, a novel Size-Translation-Rotation-Invariant Character Recognition method

Character recognition is an active field of research. Applications include point of sale systems, tablet computers, personal digital assistants (PDAs), smart phones, and military applications. Recognizing Asian characters has been pursued since 1984, and difficulties exist in Japanese due to the complexity and numbers of Kanji, Hiragana, and Katakana characters. It is further complicated by differences in size, translation, and rotation. This paper contributes an original approach to constructing feature vectors. The presented Size-Translation-Rotation-Invariant Character Recognition and Feature vector Based STRICR-FB algorithm is based on the Kohonen Winner Take All (WTA) type of unsupervised learning. The algorithm clusters a multidimensional space vectors uniquely derived from the Hiragana characters. The STRICR-FB methodology creates a neural network by design and not by training. This alleviates typical training problems like instability and no convergence. Furthermore, an upper bound degree of closeness is determined by the distance between the two closest unique feature vectors. The STRICR-FB algorithm was implemented in Matlab and uses the Image Processing Toolbox to process the images. The algorithm was tested on the MS Mincho font set. It demonstrated a recognition rate of 90% independent of size, translation, and rotation.