a new method for crater detection heather dunlop november 2, 2006
DESCRIPTION
System Overview ● Compute probability of a boundary image ● Use Hough Transform to detect circles as candidate craters ● Compute a set of features on each candidate ● Apply SVM classifier to identify craters vs. non- cratersTRANSCRIPT
A New Method for Crater Detection
Heather DunlopNovember 2, 2006
Introduction● Purpose:
– Detect as many craters as possible– With as high an accuracy as possible
System Overview● Compute probability of a boundary image● Use Hough Transform to detect circles as
candidate craters● Compute a set of features on each
candidate● Apply SVM classifier to identify craters vs.
non-craters
Boundary Image● Canny Sobel
Boundary
Probability of a Boundary● Natural image boundary detection
– Martin, Fowlkes, Malik, UC Berkeley● Brightness, texture gradients● Half-disc regions described by histograms
● Compare distributions with χ2 statistic● Combine cues to form probability of a
boundary image
r
(x,y)
Hough Transform● For lines:
– “There are an infinite number of potential lines that pass through any point, each at a different orientation. The purpose of the transform is to determine which of these theoretical lines pass through most features in an image.” -- wikipedia.org
● For circles:– Parameterize by circle center (x,y) and radius
r– Each edge point votes for possible circles by
incrementing bin in accumulator matrix– Circles with the most votes win
Detect Circles● Threshold boundary image and apply
Hough Transform
Region Features● Features that can distinguish crater from non-
crater regions● Shading● Intensity● Texture● Template● Boundary● Radius● Lighting: azimuth angle, angle of incidence
Shading Features● Mostly applicable to day images● Linear gradient due to directional lighting● Compute best fit linear gradient● Features:
– direction of gradient– strength of gradient– SSE to gradient
Crater Regions● Inside Rim Outside Whole
● Compare regions with ● Euclidean distance or • χ2 statistic
r
δ
Intensity Features● Mean intensity
● Histogram of intensities
Texture● MR8 Filter bank: Varma, Zisserman
– Edges– Bars– Spots
– Multiple orientations and scales● Convolve images with set of filters● Aggregate responses● Cluster with k-means to form textons
Texton Maps● Compute nearest texton for each image
pixel's response vector● Form texton map for image
Texture Features● Histogram of textons in region
Template Features● Mostly applicable to night images
● Crater sort of looks like this:
● Sum element-wise multiplication with image and normalize by size
Boundary Features● Sum probability of a boundary in rim
normalized by area of rim
Support Vector Machines● Linear SVM: linear separator that
maximizes the margin
● For non-linearly separable data:
http://www-kairo.csce.kyushu-u.ac.jp/~norikazu/research.en.html
http://www.cs.cmu.edu/~awm/10701/slides/svm_with_annotations.pdf
Crater vs. Non-Crater Classifier● Train an SVM classifier using features
extracted● Training data:
– ground truth craters– Hough detected circles that are not craters
● On test image, apply classifier to candidate craters to determine probability that each is a crater
Experiments● 8 day images, 8 night images● 820 craters, approx. 50 per image● Each crater 4 pixels or larger in radius
marked as ground truth● Looking for craters of minimum radius 5
pixels● Leave-out-one-image cross validation
Results: Day
Legend: False positive Detected true positive Ground truth for true positive Not detected
Results: Night
Legend: False positive Detected true positive Ground truth for true positive Not detected
False Detections
Legend: False positive Detected true positive Ground truth for true positive Not detected
Performance Metrics● Precision: fraction of detections that are
true positives rather than false positives● Recall: fraction of true positives that are
detected rather than missed
ResultsImage Precision Recall
1 0.8 0.332 0.68 0.583 0.94 0.224 0.86 0.265 0.2 0.356 0.82 0.227 1 0.388 0.6 0.419 0.72 0.37
10 1 0.4111 0.59 0.3412 0.73 0.3313 0.78 0.6214 0.89 0.7115 0.89 0.7316 0.9 0.79
DayNight
All 0.77 0.44
Conclusions● Works better on day images than night● The more training data the better
● Questions, comments, suggestions?