This paper presents a construction of a pattern recognition system invariant to translation, scale-change and rotation of patterns which are included in the same similitude transformation group (STG).
In principle, an invariant system can be achieved by constructing proper normalization on patterns subject to STG and matching them with standard patterns prepared.
From the above point of view, we propose a 3-phase method for processing in feature extraction mechanism (FEM) and feature space (FS). First, normalization for translation transformation is realized through constituting FEM with
Nth-order autocorrelation function of patterns (AFP). The calculations of AFP are performed with a high speed hardware based on mask pattern histogram method (MPHM) and patterns are mapped into a FS of 25 dimensions.
Second, normalization for scale-change transformation is carried out in the FS based on variance compression of mapping process from the FS into another FS.
Third, processing for rotation transformation is done by combination of establishing plural standard patterns of different rotative attitudes and utilizing the property of 90°-rotation-equivalent mask patterns.
Last, recognition process of an unknown pattern, after the pre-processing, matches it with standard patterns under a simple minimum distance criterion.
The results of experimental investigation, where 10 digit fonts and various kinds of their transformed patterns with respect to STG are employed, show that our system is useful.
View full abstract