Automated Solar Cavity Detection Image Processing & Pattern Recognition Athena Johnson
Outline Introduction Background Problem Statement Proposed Solution Experiments Conclusions Future Work
Introduction
background Solar Dynamics Observatory (SDO) Extreme Ultraviolet Variability Experiment (EVE) Helioseismic and Magnetic Imager (HMI) Atmospheric Imaging Assembly (AIA) 1.5 Terabytes (TB) of data per day -- I suggest completely remove any discussion about CME. -- Instead, focus on large volume of solar images using numbers and facts, and thus the need for automated detection of solar cavities. -- My understanding is there is not much previous research on solar cavities. But you do need to explain all you know about previous research on solar cavity detection. Ultimately, the audience wants to clearly know: what have been done? What is your approach?
Atmospheric Imaging Assembly (AIA) Images the Corona of the Sun Study of solar storms How they are created? How they propagate upward? How they emerge from the Sun? How magnetic fields heat the corona? -- I suggest completely remove any discussion about CME. -- Instead, focus on large volume of solar images using numbers and facts, and thus the need for automated detection of solar cavities. -- My understanding is there is not much previous research on solar cavities. But you do need to explain all you know about previous research on solar cavity detection. Ultimately, the audience wants to clearly know: what have been done? What is your approach?
SOLAR CAVITIES Currently an increase in implementations focused on Solar Cavities Off limb structures Darker elliptical structure, encompassed by lighter regions Hypothesized to be precursors to solar events Aid in establishing a predictive solar weather system -- I suggest completely remove any discussion about CME. -- Instead, focus on large volume of solar images using numbers and facts, and thus the need for automated detection of solar cavities. -- My understanding is there is not much previous research on solar cavities. But you do need to explain all you know about previous research on solar cavity detection. Ultimately, the audience wants to clearly know: what have been done? What is your approach?
SOLAR CAVITIES Labrosse, Dalla and Marshall (2010) Radial intensity profiles Support Vector Machine (SVM) Region growing Calculation of metrics Running difference on subsequent images -- I suggest completely remove any discussion about CME. -- Instead, focus on large volume of solar images using numbers and facts, and thus the need for automated detection of solar cavities. -- My understanding is there is not much previous research on solar cavities. But you do need to explain all you know about previous research on solar cavity detection. Ultimately, the audience wants to clearly know: what have been done? What is your approach?
SOLAR CAVITIES Durak and Nasraoui (2010) Exraction of principal contours Calculations on contours Adaboost -- I suggest completely remove any discussion about CME. -- Instead, focus on large volume of solar images using numbers and facts, and thus the need for automated detection of solar cavities. -- My understanding is there is not much previous research on solar cavities. But you do need to explain all you know about previous research on solar cavity detection. Ultimately, the audience wants to clearly know: what have been done? What is your approach?
Detections based on metrics Weak events missed Multiple detections Problem statement Computation times Detections based on metrics Weak events missed Multiple detections Multiple events missed Low hit rates -- show a few different types of solar cavities to help with your points.
Haar Classifier Method that Paul Viola and Michael Jones published in 2001 Four key concepts Haar-like features Integral Image Adaboosting Cascade of Classifiers
Haar-Like Features Aids in satisfying real time requirements Rectangular regions Reduces Computation Good.
Integral images Rapid computation of Haar-like features
Integral images 8+6+2+5+6+3 = 30 50-17-5+2 = 30 Original Image
adaboosting Aids in increasing the accuracy and speed Begins with uniform weights over training examples Obtain a weak classifier Update weights Weak Classifier h1(x) Like integral image, start with statement on the reason why Adaboosting is used, then explain how it works.
adaboosting Weak Classifier h2(x) Weak Classifier h3(x)
adaboosting Weak classifiers combined to form the strong classifier
Cascade of classifiers Increases the speed of detections All Haar-like features from all stages combined into a final Classifier Model Cascade of boosted classifiers with Haar-like features Again, why a cascade of classifiers is used?
Cascade of classifiers A series of classifiers are applied to every subwindow of image A positive result from the first classifier, triggers evaluation from the second classifier and so on
Initial solution -- Talk about the problems with the first model first, then the second model. -- focus on the differences when you explain the model.
Results Manually selected Training Image Sets This slide is completely out of place. If you want to show the result of the first model, show and explain the model first. Manually selected Training Image Sets Positive Samples = 100 Negative Samples = 400 ≈ 79.6% Correct detection rate was achieved
Results Missed detections in specific quadrants Detections on the Sun’s disk Overlapping detections
Proposed Solution -- Talk about the problems with the first model first, then the second model. -- focus on the differences when you explain the model.
Minimized training sets 10 Positive Images 10 Negative Images Do not just use “experiment.” Use more specific title that is in consistent with the model.
Mark regions of interest and rotate Deriving images from selected images Rotation applied to both training sets Use more specific title that is in consistent with the model.
Transform regions of interest Transformations on cavities Use more specific title that is in consistent with the model.
Preprocessing Edge Detection Hough Lines Calculate the radius Use more specific title that is in consistent with the model.
Results Derived Training Image Sets Initial image in sets = 10 Positive Samples = 3600 Negative Samples = 3600 ≈ 96% Correct detection rate was achieved I understand this 96% is the result of performance testing result. Please check out how this rate is calculated. Average of 10 runs? 20 runs? From 10-fold cross validation?
Final image with detections For each slide, you want to tell the audience something. If possible, use more specific slide title.
Conclusion Less manual work Short training times < 22 hours Wider range of detections Weak and strong cavities Fast run times < 1 second per image Higher hit rates Let the facts talk. When you say “short training time” How long exactly?
Future work Technique Improvement Reduction of False Positives Contour Detections Template Matching Customized Haar-like features
Future work Find optimal number of training sets Extract Metrics User Interface
QUESTIONS?