Classification Techniques for Quantum-Limited and Classical-Intensity Images
Abstract
Automatic classification of both quantum-limited and classical- intensity images is considered. A new discriminant vector for classical- intensity images is proposed which can also be applied to general pattern recognition problems. The proposed discriminant vector, based on convex analysis, is shown in all cases to correctly distinguish two image classes. It is the discriminant vector that maximizes the minimum separation of the values obtained by forming the inner products between it and the input images. Experiments are reported in which the discriminant vector is used to successfully distinguish first two classes, then eight classes of images. The classification problem is also considered for the case of quantum-limited input images. Quantum-limited images arise as a matter of course in applications such as night vision, low-dose electron microscopy, and radiological imaging. It can also be shown, however, that for reasons of computational efficiency it may be advantageous to use a quantum-limited imaging system as the input to the classifier. The thinner product between a quantum-limited image and a discriminant vector is shown to be a Monte Carlo estimator of the corresponding high-light-level inner product. In principle, therefore, any linear classifier can be implemented using the quantum-limited system. New solutions designed for quantum-limited images are derived using statistical decision theory. These solutions are shown experimentally to provide excellent results. The method is extended to permit classification of quantum-limited images despite in-plane rotations, and the result is demonstrated experimentally. The application of the low-light-level solutions to high-light-level situations is considered. Theses.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 01, 1989
- Accession Number
- ADA218693
Entities
People
- Miles N. Wernick
Organizations
- University of Rochester