A Lightweight Feature Descriptor Using Directional Edge Maps

Abstract

The objective of this research is to design a lightweight object detection and recognition engine that requires less space, less power and smaller budget than its PC counterparts. Specifically, we develop novel feature extraction algorithms to take ad-vantage of fixed-point arithmetic. The newly defined descriptor, known as directional edge maps (DEM), can be computed using simple addition/subtraction operations. DEMs are employed as locally invariant features to represent objects of interest. When combined with a modified AdaBoost classifier, the system can be trained to detect and recognize objects of various types. The performance of the proposed descriptor in several object recognition problems are examined and compared in terms of accuracy and efficiency against local binary descriptors (LBP) and Haar-like features.