Study of two edge detectors attending to their robustness with respect to speckle noise in SAR image

1
Study of two edge detectors attending to their
robustness with respect to speckle noise in SAR
images

• J. Martín de Nicolás-Presa, P. Jarabo-Amores, D.
  de la Mata-Moya, J.C. Nieto-Borge
• Signal Theory and Communications Department.
• Technical School. University of Alcalá.
• SPAIN

ESA-EUSC-JRC 2009 Conference on Image
Information Mining
2
Contents

• Introduction
• Mean-Shift applied to image filtering
• CFAR edge detector
• Results
• Conclusions

ESA-EUSC-JRC 2009 Conference on Image
Information Mining
3
Introduction

• Objective comparative study of two edge
  detectors attending to their robustness with
  respect to speckle noise in SAR images.
• Canny algortihm
• CFAR edge detector proposed by Touzi, Lopes and
  Bousquet.

ESA-EUSC-JRC 2009 Conference on Image
Information Mining
4
Introduction

• Usual edge detectors based on the difference
  between pixel values are inefficient when applied
  to SAR images due to speckle noise.
• These detectors need a previous filtering stage.

Detected SAR image (multilook)
Speckle noise filter (Lee or Mean-Shift)
Canny edge detector
ESA-EUSC-JRC 2009 Conference on Image
Information Mining
5
Introduction

• CFAR (Constant False Alarm Rate) detector was
  proposed in Touzi et al., 1988 as a robust
  solution when there is speckle noise.
• It takes into account the multiplicative
• model of the speckle noise.
• It takes into account the statistical
• properties of the speckle noise.
• Detection based on power rates.

Detected SAR image (multilook)
CFAR detector
ESA-EUSC-JRC 2009 Conference on Image
Information Mining
6
Introduction

• Speckle filtering techniques
• Mean and Medium filters
• Lee 1986, Frost and Kuan adaptive filters
• Wavelet-based denoising algorithms
• Mean-Shift can be used for speckle filtering.
• Mean-Shift combined spatial-range processing and
  the corresponding bandwidths allow to smooth
  different texture areas, maintaining image edges.

ESA-EUSC-JRC 2009 Conference on Image
Information Mining
7
Introduction

• Mean-Shift filtering was introduced in SAR
  imagery for shadow extraction and building
  reconstruction Cellier 2005.
• Mean-Shift filtering have been compared to Lee
  filtering for speckle noise reduction using
  Envisats ASAR images Jarabo 2009.
• Results proved that Mean-Shift can improve on the
  Lee filter in speckle noise reduction and
  textures and edges preservation.
• This work focuses on edge detection.

ESA-EUSC-JRC 2009 Conference on Image
Information Mining
8
Mean-Shift applied to image filtering

• The Mean-Shift algorithm is a non-parametric
  technique for density gradient estimation
  Fukunaga and Hostetler, 1975.
• Objective to solve without
  estimating the pdf.
• Given n points xi in Rd, the Mean-Shift is
  calculated as

ESA-EUSC-JRC 2009 Conference on Image
Information Mining
9
Mean-Shift applied to image filtering

• Application to image processing
• For each pl, a sequence zj is obtained using a
  kernel like

Spatial part
Range part
ESA-EUSC-JRC 2009 Conference on Image
Information Mining
10
Mean-Shift applied to image filtering

• For the first iteration of pixel pl
• In the j-th iteration, zj1 is calculated as

ESA-EUSC-JRC 2009 Conference on Image
Information Mining
11
Mean-Shift applied to image filtering

• In each iteration the difference is evaluated
• If lt bound or jmaximum number of
  iterations

ESA-EUSC-JRC 2009 Conference on Image
Information Mining
12
CFAR edge detector

• Given a window of NxN pixels
• Two nonoverlapping equal-sized neighborhoods on
  opposite sides of the pixel under study are
  determined.
• We compute the ratio R of the average of
• Pixel values of the two neighborhoods for a power
  image.
• Square of pixel values of the two neighborhoods
  for an amplitude image.

Neighborhood 1
Neighborhood 2
ESA-EUSC-JRC 2009 Conference on Image
Information Mining
13
CFAR edge detector

• Assuming that the pixels are independent, the
  conditional pdf of R is
• L is the number of looks.
• P1 and P2 are the average powers.
• n1 for a power image, n2 for an amplitude
  image.
• This is the pdf within a homogeneous area if
  P1P2 and across the boundary in the other cases.

ESA-EUSC-JRC 2009 Conference on Image
Information Mining
14
CFAR edge detector

• R must be bounded, so a derived ratio detector r
  lying between 0 and 1 is defined as
• r R if R1
• r R-1 if Rgt1
• The conditional pdf of the bounded ratio detector
  r is

ESA-EUSC-JRC 2009 Conference on Image
Information Mining
15
CFAR edge detector

• The contrast ratio of two homogeneous areas is
  defined as Crmax(P1/P2,P2/P1).
• Given a decision threshold T, the conditional
  probability of detection, Pd, within a boundary
  between two homogeneous regions of a contrast
  ratio Cr is
• Hence, the false alarm probability is defined as

ESA-EUSC-JRC 2009 Conference on Image
Information Mining
16
CFAR edge detector

• Pfa as a function of the threshold T

1 look
6,5 looks
ESA-EUSC-JRC 2009 Conference on Image
Information Mining
17
CFAR edge detector

• This study assumes that the neighborhoods
  contiguous to a boundary are homogeneous and that
  Pd depends on the selected direction of split.
• In order to detect more edges, the operator must
  be applied in many directions.

ESA-EUSC-JRC 2009 Conference on Image
Information Mining
18
CFAR edge detector

• Iterative method for edge detection
• Given a low Pfa and a set of m neighborhoods of
  increasing size, the threshold Ti for every
  neighborhood size is computed.
• The operator r is computed over the neighborhood,
  starting with the smallest one, along the four
  directions considered.
• The minimum ratio is assigned to the considered
  point.
• The r value is retained if it lies in the
  interval Ti-1,Ti.
• The next neighborhood size is selected and the
  process is repeated until the biggest
  neighborhood is selected.

ESA-EUSC-JRC 2009 Conference on Image
Information Mining
19
Results

• TerraSAR-X

ESA-EUSC-JRC 2009 Conference on Image
Information Mining
20
Results

• Zoomed area of the mouth of the Mississippi river
  (Atchafalaya Bay).
• Objective to detect the coastlines and the
  islands.

ESA-EUSC-JRC 2009 Conference on Image
Information Mining
21
Results

• Edges detected by the Canny detector

Without speckle filter and Canny threshold 0,25
Lee filter with 5x5 window and Canny threshold 0,4
Mean-Shift with hs8, hr0,05 and Canny threshold
0,4
ESA-EUSC-JRC 2009 Conference on Image
Information Mining
22
Results

• Edges detected by the CFAR detector

Pfa0,05 3x3 window
Pfa0,05 3x3 and 5x5 windows
Pfa0,1 3x3 window
ESA-EUSC-JRC 2009 Conference on Image
Information Mining
23
Results

• Comparison of best results

Mean-Shift with hs8, hr0,05 and Canny threshold
0,4
Pfa0,1 3x3 window
ESA-EUSC-JRC 2009 Conference on Image
Information Mining
24
Conclusions

• Two edge detectors, Canny and CFAR, have been
  applied to a SAR image.
• Canny is applied with no speckle filtering, using
  the Lee filter and using Mean-Shift.
• CFAR detector is applied directly to the original
  speckled image.
• Attending to the detection of coastlines and
  islands
• Canny detector is less robust against speckle
  noise. Its performance improves considerably when
  it is combined with Mean-Shift due to its
  aptitude of smoothing homogeneous areas while the
  edges are preserved.
• Despite not using filtering, the CFAR detector
  improves on the best results obtained with the
  Canny detector.

ESA-EUSC-JRC 2009 Conference on Image
Information Mining