Building a cascade detector and its applications in automatic target detection

Abstract

A hierarchical classifier (cascade) is proposed for target detection. In building an optimal cascade we considered three heuristics: (1) use of a frontier-following approximation, (2) controlling error rates, and (3) weighting. Simulations of synthetic data with various underlying distributions were carried out. We found that a weighting heuristic is optimal in terms of both computational complexity and error rates. We initiate a systematic comparison of several potential heuristics that can be utilized in building a hierarchical model. A range of discussions regarding the implications and the promises of cascade architecture as well as of techniques that can be integrated into this framework is provided. The optimum heuristic--weighting algorithms--was applied to an IR data set. It was found that these algorithms outperform some state-of-the-art approaches that utilize the same type of simple classifier.