Affine Normalized Invariant Feature Extraction using Multiscale Gabor Autoconvolution

The paper presents a hybrid technique for affine invariant feature extraction with the view of object recognition. The proposed technique first normalizes an input image by removing affine distortions from it and then spatially re-samples the affine normalized image across multiple scales, next the Gabor transform is computed for the resampled images over different frequencies and orientations. Finally autoconvolution is performed in the transformed domain to generate a set of 384 invariants. Experimental results conducted using four different standard datasets confirm the validity of the proposed approach. Beside this the error rates obtained in terms of invariant stability are significantly lower when compared to Fourier based MSA, which has proven itself to be better than moment invariants