Sparse Coded SIFT Feature-Based Classification for Crater Detection

Morphological classification of impact craters upto now is done through visual interpretation which suffers from high degree of subjectivity. We are proposing a classification approach to classify a given crater into crater or non-crater class. The approach has been implemented and tested on Lunar images. We have also implemented some of the existing approaches like Hough transform, template matching and supervised classification using AdaBoost and found limitations of these. Our proposed framework uses sparse coded SIFT as local features and use SVM as classifier. We have compared this with other classification approaches employing pixels as features; using SIFT as local feature without applying sparse coding; using SIFT + sparse coding but employing KNN. Since SIFT demands huge memory requirement, we suggest a two phase process in which crater candidates are identified and SIFT features are extracted only for these identified areas. This drastically reduces memory and processing requirements. On testing, our approach is found to reduce the memory and time requirements almost to 50% yet outperforming in terms of accuracy as well as robustness against noise, occlusion, different viewing angles, and various illumination effects.