Hyperspectral image unmixing.
My.PC HengBao.Desktop
Code for generating spectral samples with dirichlet distribution.
VCA implementation. testVCA.m is the corresponding test code.
Nfindr implementation. testNfindr.m is the corresponding test code.
NMF with minimum distance constraint on endmembers.
NMF with minimum volume constraint on endmembers.
NMF with L1 sparsity constraint on abundance matrix.
NMF with L1_2 sparsity constraint on abundance matrix. L1_2 sparsity constraint seems inconsistent with reconstruction error.
During the test, if constraint of sum-to-one constraint is set to be small (less than 1), L1_2 constraint will get very small abundance matrix (sum of coefficient at one pixel is far smaller than 1). However, if constraint of sum-to-one is set to me large (5~10 in synthetic data), resultant endmembers will be close to true value.
Another thing is: when the value of reconstruction error and L1_2 constraint is in the similar level, reconstruction error will first decrease, while L1_2 constraint normally keep unchanged or slightly increase. when reconstruction error go down to a value far smaller than L1_2 constraint, L1_2 constraint will be the dominant factor in the following iterations. Namely, decrease of reconstruction error and L1_2 constraint are not in the same cycles. In some situations in which constraint of sum-to-one is weak, reconstruction error will go up again during decreasing of L1_2 constraint.
test codes for algorithms.
detailed experiments.