Improving the discrimination of synthetic discriminant filters

Abstract

Synthetic Discriminant Filters (SDF) are theoretically capable of distortion invariant multiclass recognition. However a serious practical limitation is the trade-off between invariance and specificity. Three techniques were investigated for improving the ability of the SDF to perform 2-D rotation invariant recognition of an industrial component from amongst similarly shaped objects. The first technique involved extracting features from the correlation image. A trade-off was found between the number and/or complexity of features which must be extracted and the number of training set images in the SDF. The second technique was edge enhancing the SDF and/or test images. Edge enhancing the test images gave greater discrimination improvement than did edge enhancing the SDF. The third method was phase only filtering. Phase only SDFs gave a dramatic improvement in discrimination between training set targets and clutter although rejection of non training set targets increases. Increasing the number of training set image remedied this.