TSRM INVARIANT RECOGNITION IN QUALITY CONTROL

NOVÁKOVÁ, K.; KUKAL, J.

Abstract

Binary image of individual n-dimensional object is an information source for object recognition. The properties extracted from given binary image should be invariant to translation (T), scaling (S), rotation (R) and mirroring (M) of the original pattern, object, or image, respectively. There are many possibilities how to realize TSR invariant properties of n-dimensional binary images. The translation invariance can be achieved by using n-dimensional Fourier transform and amplitude spectrum, which is trivial. The rotation of original will cause rotation of Fourier spectrum. Thus the rotation invariance is based on envelopes, which are generated by rotation of Fourier spectrum. Envelope invariance to mirroring (M) is the side effect, which is useful in many applications. The resulting envelopes can be rescaled to normalized forms which are TSRM invariant. The recognition system uses TSRM invariant envelopes as non-linear preprocessing for any artificial neural network. Three types of ANN are used for final processing: OLAM, RBF, SOM. Standard PCA technique is used as referential method. The TSRM recognition system is applied to quality control in food industry based on grain classification. All the programs were realized in the Matlab environment.

Coresponding author e-mail: novakovakaterina[at]seznam[dot]cz

Session: Applications and Case Studies