Motivation and Context
A significant part of planetary data is corrupted by noise, bad pointing, exposure time or gain, i.e. beyond the point that they have much scientific meaning. The available planetary datasets include many such "bad data", which both occupy valuable scientific storage resources and create false impressions about planetary data availability for specific planetary objects or target areas.
Generic Problem Description
The development of techniques for automated assessment of planetary image quality with the minimum possible amount of a priori information.
Project Goals
- Establish image descriptors to gauge the quality of a (planetary) image.
- Develop techniques that employ the above descriptors in a two-set classifier scheme that discriminate between "high-quality" and "low-quality" (planetary) images.
- Extend the classification process in order to introduce an automatic (planetary) image quality ranking system.
- Demonstrate the application of this schema to datasets where large-scale corruption is commonplace (e.g. Viking Orbiter images).
Image Descriptors Used for Quality Assessment (in Italics the novel descriptors)
- Power Spectrum Slope
R. Liu, Z. Li and J. Jia, Image Partial Blur Detection and Classification, IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1-8, 2008. - Image Anisotropy
S. Gabarda and G. Cristobal, Blind Image Quality Assessment Through Anisotropy, Journal of the Optical Society of America, Vol. 24, No. 12, pp. B42-B51, 2007. - Edge Profile Kurtosis
J. Caviedes and S. Gurbuz, No-reference sharpness metric based on local edge kurtosis, IEEE International Conference on Image Processing (ICIP), Vol. 3, pp. 53-56, 2002. - Image Self-Similarity
- Local Contrast
- Image Pixel Pairwise Statistics
Classification and Assessment
The above six descriptors are combined into a 36-dimensional feature vector per image which forms the input of two distinct classification schemes:
- A RBF-kernel SVM that discriminates between high-quality and low-quality planetary images.
- A novel, Monte Carlo meta-classification scheme that employs several SVMs to rank planetary images between interval 0 (lowest quality image) to 4 (highest quality image).
Experimental Dataset
Planetary visual spectrum images that were acquired from Viking Orbiter missions.
- Active between 1976 and 1980, during which ~47,000 visual spectrum images were acquired.
- Due to relatively primitive technology, various types of corrupt data, which in total can reach up to 50% of the total available images.
Experimental Results
Manual Annotation: 250 "high-quality" and 250 "low-quality" images
- Leave-one-out classification results: False Acceptance Ratio 6.4%, False Rejection Ratio 9.6%, (F-Score) Correct Classification 91.98% .
- Automatic assessment of 25% of the Viking Orbiter images.
- Comparison with manual assessment (future work).
Examples
Viking Orbiter images automatically assessed to be of high-quality
Viking Orbiter images automatically assessed to be of low-quality