Automatic Target Recognition and Indexing by Non-Orthogonal Image Expansion and Data-Dependent Normalization with Implementation.
Abstract
This research is concerned with the development of a neural system for robust projective-invariant recognition of multiple targets which may be partially occluded in a cluttered background based on single gray-level images. For this purpose we have developed in the research a new method for affine-invariant iconic representation and recognition of targets using a novel set of Gabor/Fourier kernels with multi-dimensional indexing in the frequency domain. An affine-invariant representation of local image patches is extracted in the form of spectral signatures, by directly convolving the image with our novel configuration of these kernels. We achieved 100% correct recognition rates with a model library of 26 models over a wide range of viewing poses and distances (360 of rotation and tilt and 82 of slant and 4 octaves of scale). The system also maintains its 100% recognition rate in high levels of noise/clutter (up to -17 dB) and to resolution degradation (1:5 reduction). A novel method for representation and recognition of 3D Object/Targets based on 3D frequency domain representation was also developed and tested.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 20, 1997
- Accession Number
- ADA333426
Entities
People
- G. Atkin
- Jezekiel Ben-arie
Organizations
- University of Illinois at Chicago