Evaluation of early-visual processing techniques for automatic object recognition

We use a modular object recognition system as a platform to evaluate the performance of various image processing techniques. The recognition system consists of modules for image restoration, detection, segmentation, feature extraction, invariant mapping, and classification. We are developing the system to classify objects in laser radar range imagery. The stages of the system preceding the classification stage are collectively referred to as preprocessing or early-visual processing because of the analogy with biological vision. In previous work, we presented results on invariant mapping techniques and concluded that the bi-directional log- polar mapping (BLP) method gave the best performance when evaluated within the context of an object recognition system. In the present study, we employ the BLP invariance module and use similar criteria for evaluation of several candidate image restoration and feature extraction modules. We use synthetic laser radar images of four vehicles rotated to various orientations in the field of view, scaled to various ranges, and corrupted by increasing levels of sensor noise for this evaluation. This study indicates that Markov-Random-Field image restoration and features extraction based on graded edges are a combination that provides the best recognition performance, as well as robustness to noise and discretization.