基于Gentle Adaboost的多姿态人脸检测器结构研究

Detectors' Structure of Multiview Face Detection Based on Gentle Adaboost

研究在多姿态人脸检测问题中,不同检测器组织结构及其性能差异。只探讨平面外旋转±90°的姿态空间,将其划分为5类姿态。首先训练单姿态检测器,通过提取扩展Haar-like特征,采用Gentle Adaboost算法,训练得到瀑布型检测器;然后对单姿态检测器进行组合构成多姿态分类器。主要对并行级联结构、并行级联附加姿态估计、金字塔结构和决策树结构4种组合结构进行实验对比,分析其在检测速度、检测率、误检数等性能指标上的差异。

This paper studies the different organization structures of the detectors and their performance differences in multi-view face detection problems.This paper only considers view space of ±90° out-of-plane rotation,and divides it into 5 categories.Firstly,train the single-view detector by extracting the extended Haar features,using Gentle Adaboost algorithm to train a cascade detector.Then,combine the single-view detectors to form multi-view detector.The paper mainly compares four structures：parallel cascade,parallel cascade with pose estimation,pyramid structure,decision tree structure,and analyze their differences of performances in detection speed,detection rate and number of false detections.