Functions of Multiple Instances, Extended Functions of Multiple Instances and Dictionary Learning using Functions of Multiple Instances
NOTE: If you use this code, cite it: Changzhe Jiao, & Alina Zare. (2019, April 12). GatorSense/FUMI: Initial Release (Version v1.0). Zenodo. http://doi.org/10.5281/zenodo.2638304
NOTE: If cFUMI and eFUMI Algorithms are used in any publication or presentation, the following reference must be cited:
C. Jiao, A. Zare, "Functions of Multiple Instances for Learning Target Signatures," IEEE transactions on Geoscience and Remote Sensing, Vol. 53, No. 8, Aug. 2015, DOI: 10.1109/TGRS.2015.2406334
If DL-FUMI Algorithm is used in any publication or presentation, the following reference must be cited:
C. Jiao, A. Zare, "Multiple Instance Dictionary Learning using Functions of Multiple Instances," in IEEE International Conference on Pattern Recognition, Cancun, Mexico, 2016.
available: http://arxiv.org/abs/1511.02825
The command to run the cFUMI algorithm:
[E, P]=cFUMI(Inputdata,labels,parameters)
% Inputs:
% Inputdata - hyperspectral image data, can be either in data cube or linear data
% parameters - struct - parameter structure which can be set using the 'cFUMI_parameters' function, see 'cFUMI_parameters' for more explanation
% labels - binary values, the same size as input data, indicates positive data point with logical '1'
%
% Outputs:
% E - Endmembers, d by M+1, d and M account for wavelength bands and number of background endmembers,respectively
% P - Proportion Values, M+1 by N, N accounts for the total number of input data
The command to run the eFUMI algorithm:
[E, P]=eFUMI(Inputdata,labels,parameters)
% Inputs:
% Inputdata - hyperspectral image data, can be either in data cube or linear data
% parameters - struct - parameter structure which can be set using the 'eFUMI_parameters' function, see 'eFUMI_parameters' for more explanation
% labels - binary values, the same size as input data, indicates positive bags with logical '1'
%
% Outputs:
% E - Endmembers, d by M+1, d and M account for wavelength bands and number of background endmembers,respectively
% P - Proportion Values, M+1 by N, N accounts for the total number of input data
The command to run the DL-FUMI algorithm:
[E, P]=DLFUMI(Inputdata,labels,parameters)
% Inputs:
% Inputdata - MIL data, d by N, d: number of data dimension, N, total number of data
% parameters - struct - parameter structure which can be set using the 'DLFUMI_parameters' function, see 'DLFUMI_parameters' for more explanation
% labels - binary the same size as input data, indicates positive bag with logical '1'
%
% Outputs:
% E - Estimated dictionary set, d by T+M, T is number of target concepts, M accounts for the number of background endmembers
% P - Representation Values, T+M by N
Files explanation:
Latest Revision: April 2016
'...\FUMI_code_and_demo':
README.md - this file
FUMI.pdf - the referred paper
Investigation of Small Increase in eFUMI Objective Function Value during Optimization.pdf - investigation of increase phenomenon in eFUMI objective function
'...\FUMI_code_and_demo\cFUMI_code':
cFUMI.m - cFUMI (convex Functions of Multiple Instances), semi-supervised target concept learning algorithm
cFUMI_Cond_Update.m - the objective calculation function
cFUMI_E_Update.m - endmember matrix update function
cFUMI_P_Update.m - proportion matrix update function
cFUMI_parameters.m - parameters generating function used during the cFUMI algorithm, please change accordingly
cFUMI_VCA_initialize.m - cFUMI initialization function
FUMI_reshape.m - reshape 3D hyperspectral data cube into linear data
FUMI_unmix.m - fully constrained least square unmixing function
FUMI_viewresults.m - proportion map display function
normalize.m - normalization function
VCA.m - Vertex Component Analysis algorithm, see reference of the referred paper for more details
'...\FUMI_code_and_demo\DL_FUMI_code':
DLFUMI.m - DLFUMI (Dictionary Learning using Functions of Multiple Instances): semi-supervised target concept learning algorithm
DLFUMI_Cond_Update.m - the objective calculation function
DLFUMI_E_Update_individual.m - dictionary set update function, atom by atom
DLFUMI_initialize.m - DLFUMI initialization function
DLFUMI_P_Update.m - representation matrix update function
DLFUMI_parameters.m - parameters generating function used during the DLFUMI algorithm, please change accordingly
DLFUMI_Prob_Z_Update.m - probability estimate function to conduct the E-step of DLFUMI
normalize.m - normalization function
my_OMP - sparse coding function using orthogonal matching pursuit
'...\FUMI_code_and_demo\eFUMI_code':
eFUMI.m - eFUMI (extended Functions of Multiple Instances), semi-supervised target concept learning algorithm
eFUMI_Cond_Update.m - the objective calculation function
eFUMI_E_Update.m - endmember matrix update function
eFUMI_P_Update.m - proportion matrix update function
eFUMI_parameters.m - parameters generating function used during the eFUMI algorithm, please change accordingly
eFUMI_Prob_Z_Update.m - probability estimate function to conduct the E-step of eFUMI
eFUMI_VCA_initialize.m - eFUMI initialization function
FUMI_reshape.m - reshape 3D hyperspectral data cube into linear data
FUMI_unmix.m - fully constrained least square unmixing function
FUMI_viewresults.m - proportion map display function
normalize.m - normalization function
VCA.m - Vertex Component Analysis algorithm, see reference of the referred paper for more details
'...\FUMI_code_and_demo\FUMI_demos':
demo_generate_synthetic_data.m - generates different type of synthetic data used in the synthetic experiment part of the referred paper
demo_FUMI_random_data_repeat_Fig_2.m - repeats the synthetic experiment of cFUMI and eFUMI on random data, corresponds to results shown in Fig. 2 of the referred paper
demo_cFUMI_noisy_data_repeat_Fig_3_a.m - repeats the synthetic experiment of cFUMI on noisy data with SNR 10, 20, 30 and 40 dB, corresponds to results shown in Fig. 3(a) of the referred paper
demo_eFUMI_noisy_data_repeat_Fig_3_b.m - repeats the synthetic experiment of eFUMI on noisy data with SNR 10, 20, 30 and 40 dB, corresponds to results shown in Fig. 3(b) of the referred paper.
'...\FUMI_code_and_demo\gen_synthetic_data_code':
add_noise_to_dB.m - adds Gaussian white noise to synthetic data
drchrnd.m - generates random vector following Dirichlet distribution
gen_individual_LMM_point.m - generates individual synthetic data point following the linear mixing model
gen_multi_tar_mixed_data.m - generates synthetic data set following the definition of multiple instance learning problem
'...\FUMI_code_and_demo\synthetic_data':
E_truth.mat - groundtruth endmember and wavelength information used to generate synthetic data
highly_mixed_data_pt_03.mat - highly mixed synthetic data with P_t_mean=0.3; bag-level and instance-level labels
highly_mixed_data_pt_05.mat - highly mixed synthetic data with P_t_mean=0.5; bag-level and instance-level labels
highly_mixed_data_pt_07.mat - highly mixed synthetic data with P_t_mean=0.7; bag-level and instance-level labels
noisy_data_SNR_10.mat - noisy synthetic data with SNR=10 dB; bag-level and instance-level labels
noisy_data_SNR_20.mat - noisy synthetic data with SNR=20 dB; bag-level and instance-level labels
noisy_data_SNR_30.mat - noisy synthetic data with SNR=30 dB; bag-level and instance-level labels
noisy_data_SNR_40.mat - noisy synthetic data with SNR=40 dB; bag-level and instance-level labels
random_data.mat - random synthetic data; bag-level and instance-level labels
For any questions, please contact:
Alina Zare
Email Address:azare@ece.ufl.edu
University of Florida, Department of Electrical and Computer Engineering