Department of Image Processing

The experiment with recognition of ALOI images was performed on publicly available Amsterdam Library of Object Images (ALOI). We downloaded 1000 images of distinct objects and rotated each of them 35 times with the step 10°. We used the nearest neighbor interpolation to resample the images for the rotated versions of the test object. The objects are masked, so the background is zero and does not influence the moment values.

To calculate the invariants of Zernike, Chebyshev--Fourier, Appell V and Appell U moments, the image was mapped onto the unit circle. For Gaussian--Hermite moments, we used the modulation parameter σ = 0.3. In all cases, invariants up to the fourth and fifth orders were applied.
To eliminate random errors and to get higher significance, we applied a multiple cross-validation. In the first run, we used the original images as the templates and classified all rotated versions. In the second run, the images rotated by 10° were used as the templates and all others were classified. We repeated this process for all rotations up to 80° (incorporating rotations beyond the first quadrant does not make sense because the results were identical).

 The entry point is the file "experiment_ALOI_rotations_confidence.m".