An adaptive learning methodology for intelligent object detection in novel imagery data

摘要

The process of robustly identifying targets embedded in a cluttered image scene is a difficult task to accomplish. Such an application must deal with rotation, scaling, and lighting variants of the target as well as handle the varying degrees of unpredictability in the image scene itself. To assume that an object will always reside in the same background environment during the detection process as in the learning phase is an overgeneralization that is unrealistic. To address this problem, we present a technique that learns to identify targets embedded in a cluttered image scene and robustly re-trains when presented with novel imagery data. The algorithm utilizes a two-stage process in which a baseline clusteringeural network methodology is used to first recognize targets embedded in an original image data set and an adaptive clusteringeural network technique is subsequently applied as images are re-examined for novelty. We show that the algorithm developed achieves a 99% recognition rate with a 0.9% false alarm rate for previously unseen background images.