Feature-Based Recognition of Objects
Paul A. Viola · 1993
We describe a technique, called feature-based recognition (FBR), that correctly classifies images of objects under perturbation by noise, rotation and scaling. FBR uses a set of feature detectors to build a representation vector for images. The feature detectors are learned from the dataset itself. A second sys-tem, called FBR+, uses a simple neural net-work to significantly improve generalization to novel images. 1