P1252428255YoRWK

1
Classification of Underwater RGB Images with
Applications in the Study of Coral Reefs

José A. Díaz Santos, Graduate Student (UPRM),
Raúl E. Torres Muñiz, PhD. (UPRM) and Roy A.
Armstrong, PhD. (UPRM)

• Abstract
• Coral Reefs ecosystems are complex
  communities in which life diversity surpass all
  ecosystems on planet. Moreover, in the last
  decades these ecosystems have been dying
  excessively, declining the coral communities
  worldwide. The objective of this research is the
  development of a classification algorithm with
  applications in the study of coral reefs RGB
  images in order to monitor those communities
  automatically. The main problem in the
  classification is that some classes are
  overlapped (Figure .4), for that reason typical
  classifiers that used primary order statistics
  like mean and covariance cannot be used. The
  classification algorithm developing in this
  research use the Local Homogeneity Coefficient
  (LHC) algorithm 5 as first stage to find the
  different regions of interest in the image like
  corals, rocks and sand among others classes.
  Then, using different texture features like
  spectral features and statistical features the
  classification of each region will be perform.
  Finally, the performance and the results of the
  algorithm will be validated with the manual
  classification done with the Canvas software
  (Figure .3). In this research, a set of 100
  images are used to test the algorithm. At this
  moment, some classifiers as Euclidean, neural
  networks and maximum likelihood 6 failed in the
  classification, with classification percent
  around 50 (Table .1).
• Introduction
• One important tool in image analysis that
  is investigated in CenSSIS (The Center for
  Subsurface Sensing and Imaging Systems) is image
  classification that is the main objective of this
  work. In this research, an automated underwater
  vehicle (AUV) (Figure .1) performed optical
  sensing using a 12-bit 1280x1024 monochrome CCD
  camera at approximately 30-80 meters depth
  8,9. These images have low contrast, and are
  very noisy. Also, they are extremely rich in
  both spectral variability and texture (Figure
  .2).
• Figure .1 The Automated
  Underwater Vehicle 8,9.
• In image classification, the acquisition
  process is a crucial step, especially in image
  rich in texture variability. This acquisition
  process is affected by many factors like change
  in environmental conditions, noise contamination
  by the water and the sensor, varying
  illumination, change in view point of the sensor
  and geometry of the sea bed, converting this
  process in a random process. For that reason, a
  robust feature selection method is needed,
  invariant for those kinds of disturbances, indeed
  of traditional first order statistics as mean and
  covariance. At this time, texture features have
  being studying for further implementation. There
  are some textures approaches in the literature of
  image classification as statistical with the use
  of co-occurrence matrices, spectral with the use
  of different filters response like Gabor filters
  and the multiresolution approach with the use of
  the wavelet transform 1, 2, 3, 4, 6,
  7.

• Objectives
• Develop a classification algorithm based on the
  segmentation of the images.
• Evaluation and comparison of different feature
  extraction and feature selection techniques in
  order to improve the classification performance
  of the classifiers.
• Evaluate the error percent of the classification
  algorithm by the comparison of the classifiers
  results and the classification results of the
  Canvas software (Figure .3).
• Methodology
• There are 21 different classes for classification
  in the 100 images, where some of them are
  overlapped (Figure .4).

Results
Classifier Results With 100 Images Mean Percent (True/True) (False/False) Better Percent Worst Percent
Euclidean (Linear) RGB 56.70 75.30 44.20
Euclidean (Linear) HSI 39.96 64.28 23.03
ML (Quadratic) RGB 49.98 77.15 33.09
ML (Quadratic) HSI 57.34 73.75 29.34
Angle RGB 50.73 84.53 31.22
Angle HSI 38.70 64.65 23.17
Euclidean and Texture RGB 61.92 78.36 31.20
Neural Networks RGB (3 layers 27-27-21) 47.97 57.49 37.80
Table 1.Classification Results

• Conclusion
• The classification results in Table .1 show
  that the high variability in texture of these RGB
  underwater coral reefs images cannot be
  classified using first order statistics. The
  feature space is not enough for the proper
  classification of these RGB images, for that
  reason methods that can increase the parameters
  in the feature space are needed, and the texture
  features may be a possible solution to overcome
  this problem in the classification.
• Present and Future Work
• Classify the difference classes of coral
  separately, because there some classes that can
  be classify easier than others and also include
  the confusion matrices of all the classes.
• Explore different color spaces like HSI,YIQ and
  XYZ 4 among others, looking for the best
  discrimination of the classes.
• Implementation of the different texture features
  selection as the statistical approach, the
  spectral approach and the multiresolution
  approach.
• Combine these new feature selection techniques
  with spatial characteristics of the sea floor,
  for improve the classification accuracy of the
  classification algorithm.
• References
• 1 J.K Shuttleworth, A.G Todman, R.N.G
  Naguib, B.M Newman, M.K Bennett, Colour texture
  analysis using co-occurrence matrices for
  classification of colon cancer images Proc. IEEE
  CCECE. Canadian Conference on Volume 2, 12-15 May
  2002 Page(s)1134 - 1139 vol.2, 2002
• 2 Liu Xiuwen and Wang DeLiang, Texture
  classification using spectral histograms IEEE
  Transactions on Image Processing, Volume
  12, Issue 6, Page(s)661 670 June 2003.
• 3 M. Soriano, S. Marcos, Caesar Saloma,
  Image Classification of Coral Reef Components
  from Underwater Color Video MTS/IEEE Conference
  and Exhibition, Volume 2, 5-8 Nov. 2001
  Page(s)1008 - 1013 vol.2, 2001.
• 4 R.Gonzalez and R. Woods, Digital Image
  Processing, 2nd ed. New Jersey Prentice Hall,
  2002.
• 5 Rivera Maldonado Francisco J.,
  Segmentation of Underwater Multispectral Images
  with Applications in the Study of Coral Reefs,
  MS Thesis, University of Puerto Rico Mayaguez
  Campus, Department of Electrical and Computer
  Engineering, 2004.
• 6 R. O. Duda, P. E. Hart and D. G.
  Stork, Pattern Classification, 2rd ed. New York
  John Wiley Sons, Inc., 2001.
• 7 Xiaoou Tang and W.K. Stewart, Texture
  classification using wavelet packet and Fourier
  transforms,IEEE Conference Proceedings.
  'Challenges of Our Changing Global Environment'.
  Volume 1, 9-12 Page(s)387 - 396 vol.1, Oct.1995.
  
• 8 http//soundwaves.usgs.gov/2003/05/
• 9 http//www.whoi.edu/DSL/hanu/seabed/

• Figure .4 Classes
  Patterns Distributions
• The first stage of this research is the
  implementation of the Local Homogeneity
  Segmentation algorithm, for image segmentation
  5.
• Next, using the segmented resulted image (Figure
  .6) from the first stage, a square of each
  segment in the image is selected. Feature
  selection is performed in this squares to perform
  the further discrimination. The features that
  will be implemented are texture features using
  three different approaches as spectral,
  statistical and Multiresolution 1, 2,
  3,7.

Figure .5 Classification Algorithm

• The classification of each segment in the image
  is performed using the features selected with
  different classifiers as Maximum Likelihood among
  others.
• The performance of the classification algorithm
  (Figure .5) is compared with the classification
  using the Canvas Software (Figure .3).

Figure .2 Underwater original coral reef image
Figure .6 Segmented Image
Figure .3 Image classification in Canvas Program
This work was supported in part by CenSSIS, the
Center for Subsurface Sensing and Imaging
Systems, under the Engineering Research Centers
Program of the National Science Foundation
(Award Number EEC-9986821).