development of advanced segmentation-based ... ao final workshop, rome, 27-29 march 2012 project id:...
TRANSCRIPT
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
Development of advanced segmentation-based multiresolution methods for speckle reduction and texture restoration in high-resolution SAR imagery
L. Alparone°, F. Argenti°, T. Bianchi°, A. Lapini°, B. Aiazzi*, S. Baronti*,
M. Abbate+, C. D’Elia+, S. Ruscino+
°Department of Electronics & Telecommunications, University of Florence, Florence, Italy*Institute of Applied Physics “Nello Carrara” (IFAC-CNR), Florence, Italy
+DAEIMI, University of Cassino, Cassino (FR), Italy
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
Scenario and motivations Problem statement
• Undecimated Discrete Wavelet Transform (UDWT)• Bayesian estimation: PDF modeling in UDWT domain
Proposed method for SAR image despeckling in UDWT domain• MAP filtering with GG and L-G priors• Segmentation-based modeling and estimation
Results on COSMO-SkyMed SAR data- sensitiveness to speckle correlation (1-look and multi-look products)- sensitiveness to resampling of the data- sensitiveness to type of estimator (GG and L-G)
Quality assessment of despeckled images• Use of a novel feature to detect structural changes between two SAR images• Comparisons between the segmented GG-MAP filter and Gamma-MAP filters
Presentation Outline
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
Scenario and Motivations
• The all-weather acquisition capability of SAR systems, either spaceborne or not, is hampered by the presence of the noise peculiar of coherent systems named speckle.
• Speckle severely undermines the possibility of performing automated analysis in general, and temporal change analysis in particular.
• A preliminary processing of (detected) SAR images aimed at speckle reduction is beneficial for a number of applications but such a preprocessing should be carefully designed to avoid spoiling useful information:
• local mean of RCS;
• point targets;
• linear features;
• texture.
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
Problem Statement
• Under a perspective of statistical signal processing, despeckling aims at performing an estimation of the radar reflectivity, based on the available speckled image.
• Such an estimation may be approached as a Bayesian estimation as the conditional expectation of the noise-free reflectivity to the noisy SAR image
• minimum mean square error (MMSE) estimation (conditional mean),
• minimum mean absolute error (MMAE) estimation (conditional median),
• maximum a-posteriori probability (MAP) estimation (conditional mode).
• Under Gaussianity assumptions, all estimators yield identical solutions: linear MMSE (LMMSE) estimation or pseudo-Wiener filtering (for non stationary signals).
• The MAP approach, is generally more powerful, even though more crucial and computationally more complex.
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
Multiresolution Despeckling
• A promising approach to image denoising in general, and to despeckling in particular, consists of carrying out estimation in a transformed domain obtained via multiresolution analysis (MRA).
• The undecimated discrete wavelet transform (UDWT), a.k.a. stationary wavelet transform (SWT), exhibits the shift-invariance property and is particularly suitable for most of image processing tasks (e.g., denoising, fusion, etc.) with the exception of data compression.
• Following a trend established in the literature during the last decade, Bayesian estimation has been accomplished in the UDWT domain, thereby exploiting its scale-space varying structure.
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
Stationary Wavelet Transform 1/2
SWT
Equivalent filters of SWT
DWT
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
Stationary Wavelet Transform 2/2
DWT
SWT
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
Despeckling in UDWT Domain
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
Speckle Model (1 of 2)
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
Speckle Model (2 of 2)
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
Bayesian Estimators
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
Statistical Modeling in UDWT Domain (1 of 3)
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
Statistical Modeling in UDWT Domain (2 of 3)
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
Statistical Modeling in UDWT Domain (3 of 3)
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
Filters for G-G Model
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
Filters for L-G Model (1 of 2)
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
Filters for L-G Model (2 of 2)
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
Filters for GG-GG Model
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
Results with Simulated Images
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
MAP Despeckling in SWT Domain (1 of 2)Argenti, Bianchi & Alparone, IEEE TIP, 2006
Conjecture: SWT coefficients obey to a Generalized Gaussian law with space-varying parameters
MAP equation for GG PDFGG parameters estimation:• Standard deviation: same as for ML estimation;• Shape factor: computed from 2nd- and 4th-order moments of empirical distributions of SWT coefficients.• SWT coefficients in MAP equation obtained by LTI filtering reflectivity and speckle; • moments computed from cumulants: cumulants of LTI-filtered process obtained by LTI filtering cumulants of
input.
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
MAP Despeckling in SWT Domain (2 of 2)Bianchi, Argenti & Alparone, IEEE TGARS, 2008
The MAP solution in SWT domain has been specialised to SAR imagery in the following manner:
• Every SWT subband is segmented into statistically homogeneous segments and one shape factor of GG PDF is estimated for each segment (variance is estimated for each point, same as for LMMSE filtering).
• This solution allows to handle scene heterogeneity as imaged by SAR system.
• The concept of heterogeneity is crucial for high spatial resolution SAR: the circular Gaussian model of complex reflectivity may no longer hold when the resolution cell does not contain a large number of independent scatterers.
• Thus, the speckle model changes with the degree of texture.
• Point targets are preprocessed, i.e. detected in the spatial image and removed (clipped) before wavelet processing, to be restored in value after despeckling.
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
TS-MRF Segmentation (1 of 2)
• Image segmentation has been found to be useful for MAP despeckling.
• Unlike the earlier work in which wavelet planes were individually segmented based on the energy of their coefficient, in this work, a segmentation map is calculated for the whole image and superimposed to all wavelet planes.
• Problem: image segmentation algorithms exploit image models of local homogeneity and SAR images are corrupted by speckle noise.
• Traditional image segmentation algorithms may be inadequate to SAR images.
•Information-Theoretic Heterogeneity Features [Aiazzi et al., IEEE TGARS, 2005] are calculated from the original SAR image.
• Such maps of features are segmented by using a tree-structured algorithm based on Markov random fields (TS-MRF) [C. D’Elia, G. Poggi and G. Scarpa, IEEE TIP, 2003].
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
Feature Map
TS-MRF Segmentation (2 of 2)
Segmentation Map
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
CSK Stripmap 1-look 1024 x 1024 (from 1A-bal.)
Original: Measured ENL=0.91 GG-MAP-SEG: Measured ENL = 18
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
CSK Stripmap 1-look 1024 x 1024 (from 1A-unb.)
Original: Measured ENL = 0.97 GG-MAP-SEG: Measured ENL = 100
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
CSK Stripmap 4-look 512 x 512 (from 1A-bal.)
Original: Theoretically calculated ENL = 2.20 GG-MAP-SEG: Measured ENL = 85
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
CSK Stripmap 4-look 512 x 512 (from 1A-unb.)
Original: Measured ENL = 3.10 GG-MAP-SEG: Measured ENL = 210
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
Effects of Resampling on Despeckling
GG-MAP-SEG of 4-look from 1A: ENL = 85 GG-MAP-SEG of 4-look from 1B: ENL = 75
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
Comparison of Estimators: GG-MAP-S vs. LG-MAP-SLG-MAP-SEG: ENL = 170 GG-MAP-SEG: ENL = 85
LG-MAP-SEG too aggressive on textureHOWEVER
LG-MAP 10 times faster than GG-MAP!
Original 4-L 300 x 512 from 1A-b data product
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
Comparisons w/ State-o-t-Art: 4-L, 512 x 512, 1A-b
Gamma-MAP filter (NEST implementation): ENL = 45 GG-MAP-SEG: ENL = 85
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
Map of Feature Showing Filtering Inaccuracies
Gamma-MAP GG-MAP-S
COSMO-SkyMed AO Final Workshop, Rome, 27-29 March 2012 Project ID: 2293
• A computational approach to despeckle SAR images based on segmented MAP filtering in UDWT domain.
• Optimization of the method on COSMO-SkyMed data products.• Multilooking mitigates the strong spatial correlation of speckle in single-look data• Influence of resampling on MAP filtering performances is little relevant.• Cost-benefits tradeoff of PDF models for MAP filtering (L-G 10 times faster than
GG: 9 s vs. 93 s for a 512 x 512 image and Matlab implementations).• The final version can switch different estimators (L-G-MAP, GG-MAP) according
to the degree of texture of individual segments.• Comparisons more than favorable with classical Gamma-MAP filter.• Developments of statistical methods to evaluate the quality of despeckling on true
SAR images without reference originals.• Blind whitening of SLC data to achieve uncorrelated 1-look intensity speckle.
Conclusions and Developments