Evaluation of the Informativeness of Multi-order Image Transforms

We studied the informativeness of image features extracted from different lengths of image transform chains for the purpose of image classification. Image features were extracted from the raw images, image transforms, and second, third and fourth order of compound image transforms. The transforms used in this study are Fourier, Chebyshev, and Wavelet (symlet 5) transform. Experimental results show that image features extracted from first and second order of compound image transforms can in some cases be more informative than the image features extracted from the raw pixels, and can significantly contribute to the classification accuracy. However, chains of transforms longer than two do not improve the classification accuracy of the image datasets used in this study.