Deep Convolutional Neural Networks Based Single Image Super-Resolution And Classification For Crater Detection
AuthorGohari, Ebrahim E.
Computer Science and Engineering
AltmetricsView Usage Statistics
Craters are among the most abundant features on the surface of many planets with great importance for planetary scientists. Craters reveal chronology information about planets and may be used for autonomous spacecraft navigation and landing. Typically, crater detection involves two main phases: (i) hypothesis generation (HG) and (ii) hypothesis verification (HV). During HG, potential crater locations are hypothesized using unsupervised algorithms. The validity of the hypothesized crater locations is then tested in a HV step using sophisticated supervised algorithms. Although many research efforts have been carried out in the field of crater detection, existing crater detection algorithms (CDAs) are only helpful in a rather limited number of applications due to several weaknesses. For instance, many CDAs make strong assumptions about the visual characteristics of craters in the HG step, leading to missed detections. Moreover, inaccuracies in the HG step lead to poor verification performance. Detection of small craters has also not been addressed sufficiently despite their importance; for example, finer spatial resolution of the geologic stratigraphy of planetary surfaces can only be obtained from the statistics of smaller craters. This is mainly because although small craters are abundant on planetary surfaces, they are challenging to label. The lack of ground truth for small craters makes it problematic to train highly accurate hypothesis verification algorithms.In the first part of this work, we investigate the performance of HG and HV separately. Our goal is to better understand how inaccuracies in the HG step affect the HV step. In this context, we evaluate the performance of two commonly used in the literature HG approaches based on highlight and shadow region detection and Hough Transform. Then, we propose two novel HG algorithms based on interest point detection and convex grouping and compare their performance with previous methods. To deal with different size craters, we employ a multiscale HG approach. We also investigate the robustness of an extensive set of HV algorithms in the presence of inaccuracies in the HG step and propose a mechanism to systematically augment the training data with miss-localized craters to improve HV performance in the presence of such inaccuracies. Based on our HV experiments, we chose Convolutional Neural Networks (CNN) to further investigate how various HG algorithms affect the HV step in order to find an optimal combination for Lunar crater detection.In the second part of this work, we address the problem of small crater detection by proposing a novel crater-specific Single Image Super Resolution (SISR) algorithm based on Deep CNNs. The key idea is the introduction of Crater Detection Loss (CDL) during the training process of the SISR network, to improve the super resolution performance on small craters. We then approach the problem of small crater detection by first increasing the resolution of the input image and then applying the original CDAs, trained on large craters, to detect small craters. Using our SISR CNN, HG based on interest point detection, and HV using CNNs, we can effectively detect small craters without having to retrain the HV algorithm by augmenting the training set with small crater examples. Our results show that we can detect craters as small as 7 meters in diameter; to our knowledge, this is the smallest target size for any CDA reported in the literature.