Spectral Indexing for hyperspectral image CBIR

1
Spectral Indexing for hyperspectral image CBIR

• Orlando Maldonado, David Vicente, Miguel A.
  Veganzones, Manuel Graña
• Dept. CCIA, UPV/EHU, San Sebastian, Spain

2
Outline

• Content Based Image Retrieval (CBIR)
• Features for hyperspectral CBIR
• Spectral unmixing
• Associative Morphological Memories
• Similarity distance between images
• Experimental results and conclusions

3
Content Based Image Retrieval (CBIR)

• Search and access in databases of images based on
  features computed from the images
• Computer vision and image processing
• Image indices are numerical feature vectors
• Image similarity distance in feature space
• Query given by image samples

4
CBIR

• Typical features of 2D images
• Shape contours
• PCA, ICA dimensión reduction
• Color regions
• High level object recognition
• Typical distances
• Euclidean, Mahalanobis
• Correlation between images
• Image registration (medical images)

5
CBIR hyperspectral images

• Conventional CBIR does not apply
• High dimensional pixels --gt band selection
  previous to any processing
• Using spectral information for indexing
• Our proposal
• Using image endmembers as indices

6
(No Transcript)
7
(No Transcript)
8
Features for hyperspectral CBIR

• Linear Spectral Unmixing (LSU)
• Given a set of endmembers compute
• abundance coefficients the fractional
  contribution of endmembers.
• Endmembers
• Selected by experts
• Induced from the data

9
Featurse for hyperspectral CBIR

• Spectral mixing
• Measured pixel spectra are the result of the
  observation of a physical mixture of materials
  characterized by their spectra.
• Linear models correspond to several land covers
  aggregated in a pixel due to lack of spatial
  resolution.
• Non linear models correspond to intimate
  mixtures.
• Endmembers spectra of the elementary materials.

10
Linear spectral mixing
11
Features for hyperspectral CBIR

• Spectral unmixing
• It is the computation of the abundance image
• Fractional contribution of each endmember to the
  observed pixel spectrum.
• For linear mixtures, it corresponds to the
  inversion of the linear model.
• Abundance coefficients desired properties
• Additivity
• Non negativity

12
Features for hyperspectral CBIR

• Endmember induction algorithm
• Morphological Independence
• Endmembers correspond to vertices of the convex
  hull of the data
• Associative Morphological Memories
• Useful to detect the morphological independence
  condition

13
Morphological Associative Memories

• Morphological Neural Networks are constructed
  isomorphic to conventional Neural Networks
• Weights are additive instead of multiplicative
• Neuron excitation is given by a max/min operator
  instead of addition

14
Morphological independence
x
x
x
x
x
x
x
x
x
15
Morphological independence and AMM

• For binary patterns, AMM are selectively
  sensitive to morphological independence.
• W memories are able to store and recall patterns
  corrupted by erosive noise. Non perfect recall
  implies dilative independence relative to the
  stored patterns.
• M perform dually and allow detection of erosive
  independence.
• Endmembers are (usually) independent in either
  erosive or dilative sense. AMM recall performs as
  the detector.

16
Detection of morphologically independent pixels
Initialization
New endmember
Next pixel
17
Example of comparative results of endmembers
extractionabundance images from Salinas A
18
Similarity measure
19
Similarity measure
20
Similarity measure

• Deals with spectral information
• Works with different number of endmembers
• Works with asymmetric situations
• One subset of endmembers of one image very close
  to the whole set of endmembers of the other image.

21
Experimental validation

• Database of simulated hyperspectral images
• USGS AVIRIS ground truth spectra
• Random Gaussian field ground truth abundances
• The validation experiment
• Most similar images agreement
• Ground truth endmembers
• AMM induced endmembers

22
Experiments

• Number of ground truth spectra 2 to 5
• Number of images per number of ground truth
  spectra 100
• Image size 256x256 pixels

23
Experimental validation
24
(No Transcript)
25
(No Transcript)
26
Conclussions

• We propose endmembers as features for
  hyperspectral image CBIR
• We define a measure of similarity
• Experimental validation on simulated images is
  encouraging.
• Further work must address the introduction of
  spatial information from the abundance images in
  the similarity measure.
• Need of ancilliary information.