Comparing Sampling Methods in Faster Computation of Non-Negative Tensor Factorization of Spectral Images

Computation of the non-negative tensor factorization of a spectral image is very time-consuming. The computational complexity depends on the number of bases, i.e. the rank of the factorization, and on the dimensions of the spectral image. In this study we propose sampling methods for the preprocessing phase which enables a faster way to compute the non-negative tensor factorization (NTF). In the preprocessing both sampling and interpolation are applied to the original data. Three approaches are compared: direct sub-sampling, integer wavelet transform, and spectral smoothing. The experiments indicate that the preprocessing can remarkable reduce the time needed for NTF. From the approaches, the integer wavelet transform shows the best performance in computational and quality senses. The computational load from the direct subsampling is the lowest for one iteration, the spectral smoothing is computationally heaviest.