Segmentation of underwater lands images.

1
Segmentation of Sea-bed Images.
Josepha UNIA Ecole Centrale de Lyon

2
Summary

• Introduction
• Sensing
• Image segmentation
• Vector quantization
• Preliminary results
• Perspectives

3
Introduction

• Marine application Mapping of living/dead
  Maerl.

• Provide the information necessary for a safe and
  sustainable exploitation of underwater resources.
• Develop and implement mathematical models, and
  sensing and guidance techniques

4
Sensing (WP3, subtask 3.1)

• Design signal processing algorithms for
• on-line extraction of features for platform
  guidance
• Segmentation of video images
• video mosaicing
• using contour information to correct (effects of)
  positioning errors

5
Image segmentation
Segmentation
Contour-extraction
Image
ROV Control
6
Vector quantization
Dictionary

• What size for dictionary ?

• What words ?

• What size for window (vector) ?

• What resemblance measure ?

7
Algorithm Outline

• Dictionary learning
• Initialize dictionary (centroid of all
  vectors - windows)
• Split until desired dictionary size is reached
• PARTITIONING make best attribution of each
  vector to current dictionary
• CORRECT DICTIONARY to get minimal distortion
  (for current partioning)

8
Preliminary results (1)
Dic
2
4
16
32
Window
3 x 3
10 x 10
30 x 30
9
Preliminary results (2)
Frame 01
Frame 02
Frame 03
Frame 04
2 CODEWORDS - WINDOW 10x10
F_out 02
F_out 01
F_out 03
F_out 04
10
Preliminary results (3)
Frame 01
Frame 02
Frame 03
Frame 04
16 CODEWORDS - WINDOW 10x10
F_out 02
F_out 01
F_out 03
F_out 04
11
Lower resolution image
What resolution ?
Window 10x10 2 clusters
12

Perspectives

• Learning
• representation of classes (coverage with enough
  vectors)
• rotation invariance