Upload
julie-casey
View
217
Download
0
Embed Size (px)
Citation preview
Content Analysis and Restoration of Marine Video
Student :- Ken Sooknanan
Supervisor :- Professor Naomi Harte in collaboration with Prof. Jim Wilson
(Dept. of Zoology), the Marine Institute Ireland, and Prof. Anil Kokaram (Google)
Goals
1) Content Analysis:-
(a) Identify Nephrop burrow complexes
(b) Identify when major changes along the seabed occur.
(c) Rocky (d) Muddy Video 2) Restoration:-
(a) Improve Visibility by correcting illumination degradations due to light source.
(a) Complex (b) Nephrop Video
Light Footprint
Improving Visibility
(2) S, c estimated with ratio of pt. correspondences
(1) Degradations modelled as mixture
optimized alternately in Bayesian Framework
G(x) )(5.3 2
)( xrexI
r ≤ rf)(xIr > rf
= );()()(2 cxScxxr T
s1 s2
s2 s3S =
cx cy
; c = )(xI )(5.3 2
)( xrexI G(x)
21
1
22
2
5.3
5.3
reI
reI
2
1
G
G 21
22
1
2 5.3ln rrG
G
cn+1=arg max p(c|G1,G2,Sn);
Sn+1=arg max p(S|G1,G2,cn+1);
(3) Footprint estimated as region where largest
percentage increase in degradation occurred G1(r1)
G2(r2)(c) A(rx) = (d) A(rx) > μ. Split into
regions, footprint (red)
Frame 2
G2
c
r2G1
c
r1
Frame 1
Results
(a) Original with
footprint in blue
(b) Our Result (c) Seon Joo
Original Result
(d) Seon Joo Result
with our radii est.
IMAGES :-
VIDEO :-
[1] Seon Joo, et al.., “Robust radiometric calibration and vignetting correction," Pattern Analysis and Machine Intelligence, 2008
(a) Sample Video Frame
(b) Generated Mosaic
(using 1st 50 Frames)
Identifying Nephrop Clusters
• Exploring the use of a Mosaic, as it:-
- Eliminates tracking
- Fixes geometric distortion.
- Computations reduced to a single image
- Provides a wide view of the seabed.
- Easy to spot clusters.
• Problem broken up into two parts:-
- identifying all burrows
- clustering
• Currently working on identification part
Algorithm overview
- Cascade Classifier,
p(f) > (T =0.5) SHAPE
(1) Highlight all dark regions
- DOG
4) Classify based on shape, colour and shading features, f,
2) Segment and label DOG – Intensity based (Bayesian framework - ICM).
3) Identify and Split multiple burrows regions – shape modelling with GMMs
(a) Original (b) Segmentation (c) Est. GMM (d) Split Region
COLOUR SHADDING
25.0
)(
f
efp
Initial Results
(a) Original Mosaic (b) Identified Burrows
Exp. Video Results Mosaic Results
GroundTruth
% CorrectDetected
GroundTruth
Recall %
Precision %
1 50 90 50 90.0 84.0
2 63 85.7 60 82.5 88.3
3 84 88.1 83 88.1 85.5
Results obtained from 3 Video
Sequences (10 min., 15000
Frames) Compared with
Ground Truth from:-
(1) Video (existing method)
(2) Mosaic
Accomplishments
(1) Gave a presentation in Google, San Francisco on 23rd Jan. 2012
(2) Published a paper [1] in SPIE conference, San Francisco Jan. 2012
(3) Submitted a paper [2] on Burrow Detection algorithm in ICIP 2012.
(4) Submitted an abstract [3] on Mosaicing Algorithm in Oceans conference 2012.
[1] Sooknanan, et al.., “Improving Underwater Visibility using Vignetting correction," in proceedings of SPIE, 2012[2] “A Bayesian Framework for Detection of Nephrop Burrows for Seabed Video Analysis,” ICIP 2012 [3] “A Bayesian Framework for Mosaic Creation of the Seabed from Underwater Video For Nephrop Burrow Detection”, Oceans, 2012
Year 1:- Attended and passed 3 courses to gain the necessary 15-credits
Year 2:- Wrote year-1 transfer report and passed viva to enter PhD roster.
Year 3:-
(5) Established good collaboration with the Marine Institute Galway
- keep in contact with Jennifer Doyle approximately once every month.
Future Work(1) Present work in Study Group on Nephrops Survey (SGNEPS) Meeting in
Italy in 8th March 2012.
(2) Do more testing with Burrow detection algorithm
(3) Move onto the Burrow clustering part of the problem
(4) Move onto detecting when major changes in seabed type occur
(5) Write papers on (3) and (4)
(6) Write up Thesis.