Primitive Feature Extraction via a Combined ICAWavelet Method

1
Primitive Feature Extraction via a Combined
ICA-Wavelet Method

• Vijay Shah, Nick Younan, Surya Durbha, and Roger
  King
• Department of Electrical and Computer Engineering
• GeoResource Institute
• Mississippi State University,
• Mississippi State, MS 39762, USA
• IIM 2008, Frascati, Italy
• March 4-6, 2008

2
Overview

• Background
• Data Transformation
• ICA-Wavelet Feature Extraction for Image
  Information Mining
• Methodology
• Coarse Image segmentation
• Region Identification
• Experimental Results
• Summary

3
Background

• Feature extraction and data transformation are
  important in any image information system
• Primitive features are extracted based on color,
  texture, and shape within a region of interest.
• Wavelet transformation has been effectively used
  for feature extraction in many applications.
• It reduces the complexity and dimensionality of
  the extracted features to expedite the image
  retrieval in EO data archives.

4
Data Transformation

• Data transformation is used to improve the
  quality of knowledge discovery in an image.
• Spectral transformation can take place in many
  different forms
• Applying arithmetic operation (,-,) on a
  feature set
• Combining features that are correlated
• Transformation along the spectral axis to obtain
  a new set features
• Linear transformation
• Nonlinear transformation

5
Common Methods

• HSV
• RGB space is nonlinearly converted to the HSV
  (Hue color type, Saturation color purity, and
  Value color intensity) space to make color
  components perceptually independent and uniform.
• Fails to capture the complete spectral pattern
  an important characteristic available in remote
  sensing imagery
• HSV space is not statistically independent and
  uncorrelated
• PCA
• Linear transformation of the RGB space.
• Not optimized for class separation

6
Current I3KR System Intelligent Interactive
Image Knowledge Retrieval
Color LAB Space Texture Features from
L-component Co-occurrence matrix Uniformity,
Entropy, First Order Element, Maximum
Probability, First Order Inverse
Element Primitive length matrix - gray level
uniformity, long primitive emphasis, short
primitive emphasis, uniformity, and primitive
percentage
7
Methodology
Data Transformation for Feature Extraction
Spectral Transformation (ICA)
Raw Imagery
Spatial-Transformation (Wavelet-Transform)
Image Segmentation
Clustering
Region Feature Extraction
8
Feature Extraction

• Component energies and Cross-correlation energies
  of the coarse scale wavelet coefficients are
  considered robust to capture color and texture
  information.
• For the l-level DWT, a n-D vector for the jth
  wavelet coefficient of B sub-band, is obtained,
  where n number of independent components.
• Appropriate wavelet decomposition level is
  selected based on the frequency content of the
  image.

9
Decomposition Level Selection

• Steps
• Calculate the Fourier transform of the image.
• Retain the frequencies whose energy is greater
  than 1 of the main peak, as the frequency
  components of the image.
• Calculate the total energy E of the spectrum.
• Calculate the total energy Ei in sub-band (i),
  over the region of with i 1,2, , N.
• If Ei/E lt threshold, increase i and repeat the
  previous step, else return (i-1) as the
  appropriate level of decomposition for the image.
  The threshold value is typically chosen to be
  close to 0 to guarantee that there is no loss in
  the frequency components of an image, i.e. higher
  threshold values will disregard low magnitude
  frequency components as being significant to the
  image content.

10
Independent Component Analysis

• Variant of Principal Component Analysis (PCA)
• Seeks those directions in the feature space that
  are statistically independent and uncorrelated
• Uses higher order statistics to determine the
  mixing matrix

11
Clustering and Object Identification

• Clustering K-means and Kernel K-means
• provide relative scalability and very efficient
  processing for very large datasets.
• Eliminates the use of cross-correlation energies
  as features (Kernel K-means), i.e., feature
  reduction during the pre-processing stage of ICA
• Object Identification - SVM
• Proven to be successful for many applications of
  nonlinear classification and function estimation
• Trained to find the global optimum
• Hyperplane that separates the two classes with
  maximum margin for linear case
• For non-linear cases, the feature space is mapped
  to higher dimension to separate two classes

12
Segmentation Evaluation

• Index I
• Maximized for when correct number of cluster are
  found
• where,
• U(X) ukjKxn is a partition matrix of the data

• Silhouette Coefficient
• Average SC provides information on cluster
  separation, and for each point given by
• where dissimilarity of point i to other pt. in
  clusters

• J-value
• Minimized for good segmentation algorithm, given
  by
• where,
• z (x,y) represent the 2-D vector image pixel
  position of the classified image z?Z
• Z is divided into C classes

13
Experimental Results

• Data Sets
• false color Landsat 7 ETM scenes
• 512 x 512 pixels each
• 30m x 30m spatial resolution.
• 4 classes - water bodies are generally dark color
  objects with smooth texture, agricultural land
  with healthy vegetation is dark pinkish-red with
  smoother texture, forest is dark red with coarse
  texture, and fallow land is yellowish-gray in
  color.

Image 1
Image 3
Image 2
14
Experiment 1 -Comparison of Different Spectral
Transformation Methods
15
Experiment 2 Spectral and Spatial Transformation
Order
Spatial-Spectral transformation order does not
matter for estimating the number of clusters when
using ICA spectral transformation
J-value reduces if the spectral transformation is
performed after taking 2D-DWT
16
Segmentation -Visual Comparison
Image 1
(a) ICA-spectral Xformation and 2D-DWT
(b) 2D-DWT and ICA-spectral Xformation
(c) JSEG
(d) HSV spectral Xformation
(e) PCA spectral Xformation
(f) No spectral Xformation
17
Experiment 3 - Clustering Approaches Comparison

• Experiment conducted on pixels from classes
• water, agricultural land, forest, and fallow land
  
• 100 samples from each class
• Total iteration is set to 100 and the number of
  replication set is 10
• Number of clusters for both algorithms is varied
  from 2 to 5
• Gaussian RBF kernel with s 1 for kernel k-means
  algorithm

18
Error Evaluation

• Error calculated based on the improper cluster
  assignment of the sample
• The omission error is defined as excluding a
  sample that should have been included in the
  cluster.
• The commission error is defined as including a
  sample in a cluster when it should have been
  excluded

19
Results - Experiment 3
k-means
Kernel k-means
20
Experimental 4 Region Feature Extraction
Comparison

• Data
• 150 region sample of each class from the Landsat7
  ETM image archive
• Total of four classes Water Fallow Land
  Agricultural Land, and Forest Area
• Used leave-one-out method for training and
  testing purpose

21
Results Experiment 4
Using the Haar Mother Wavelet
Using the rbio3.1 Mother Wavelet
22
Data - Overall System Performance

• Image Archive -
• 400 False color Landsat 7 ETM scenes
• images of size 128 x 128 pixels, 256 x 256
  pixels, and 512 x 512 pixels
• Image Resolution 28.5 m x 28.5 m for MS image
  and 14.25 x 14.25 m for Pan image
• Number of bands used 3
• Four major regions water bodies (228 images),
  agricultural land (227 images), fallow land (261
  images), and forest (224 images)

23
Results - Overall System Performance
24
Summary

• Features obtained by ICA transformation provide
  reliable segmentation compared to other
  transformation approaches.
• Choice of the order between the spectral and
  spatial transform can quantitatively affect the
  image segmentation results.
• For the ICA-spectral transformation, the
  estimated number of clusters in an image mostly
  remains the same.

25
Future Work

• Better Multiresolution Approaches
• Beyond wavelets
• Adaptive Filter Design for Multiresolution
  Approaches
• Improvement to Image Segmentation
• Redundant Approach
• Shape Feature Extraction

26
Questions?