COMPUTERIZED RECOGNITION OF SOLAR CAVITIES

Abstract

The tremendous growth in the volume of solar image data sets calls for automated feature detection mechanisms. In this thesis, an Extended Adaptive Learning Model based on Haar classifiers is proposed and implemented for automated detection of solar cavities. The model consists of a few steps. Firstly, it builds a diverse image collection for both the positive and negative training sets via image rotation and transformation. Then, the model performs Haar training and builds the classification model, in this step, simple rectangular Haar-like features are used to describe solar cavities. To facilitate efficient computation of Haar-like features, the concept of integral images is employed. Due to large number of Haar-features available, a boosted machine-learning method is performed to select features that best describe various solar cavities and generate boosted classifiers. After that, the classifier model consisting of a cascade of boosted classifiers is constructed. Lastly, an automated detector based on the classification model is used to detect solar cavities in raw solar images. To reduce false positives in the detection, the detector applies both image preprocessing techniques, such as Canny edge detection and Hough transform, and image post-processing techniques, such as duplicate detection and removal. Experiment results with Solar Dynamics Observatory (SDO) data showed that the proposed method was able to achieve an approximate hit rate of 96% and the average processing time for a single image was about I second. The proposed Extended Adaptive Learning model combines diverse image processing and pattern recognition methodologies and the implementation had a high detection rate and was able to detect a wide range of solar cavities. The study reported here has great potential to enable scientists to detect solar cavities accurately and efficiently, therefore, benefitting solar events research.