Face feature extraction | CBCL | NMF | β-divergence | Euclidean | KL | IS | mult update algo
Part-based Face feature extraction with Non-negative Matrix Factorization with different β-divergences (Euclidean: β=2, KL: β=1, IS: β=0) as loss functions, using multiplicative update algorithm - an implementation in python (using MIT CBCL faces dataset)
the generalized updates for different β-divergences (KL, Itakura) can be used to update W and H iteratively
Both W, H need to be initialized (unlike ALS algorithms where only W needs to nr initialized). Need to ensure W, H are nonnegative (e.g. with max(0, W))
For PCA, the faces are mean centered with X = (I - 11^T/n).X
As can be seen, the NMF basis images are already sparse and part-based than their PCA basis counterparts. Here only a small number of iterations (100) was used to update the matrices, if instead large number of iterations are used (e.g., ~10k) the basis images will be very sparse.
#imageprocessing #imageprocessingpython #python #machinelearning #algorithm #optimization
Watch video Face feature extraction | CBCL | NMF | β-divergence | Euclidean | KL | IS | mult update algo online without registration, duration hours minute second in high quality. This video was added by user Image Processing, CV, ML, DL & AI Projects 12 April 2023, don't forget to share it with your friends and acquaintances, it has been viewed on our site 7 once and liked it people.