人脸检测的继承式集成学习方法

基于"遗传+变异"模式,提出继承式集成学习方法框架,它可以训练出四种不同形式的层叠分类器。除了基于"无遗传"模式的基本层叠分类器与基于"全部遗传"模式的嵌入式层叠分类器两种传统方法之外,还有基于"部分遗传+部分变异"模式的特征继承层叠分类器与弱分类器继承层叠分类器。虽然后两种层叠分类器都有一定的继承代价,但是其拟合性更好,可以更好地均衡收敛速度和扩展性能,其综合性能优于传统方法。基于RAB、GAB算法与LUT弱分类器的正面直立人脸检测实验结果表明了新的继承式集成学习方法的有效性。

Inherited boosting learning for face detection

The framework of the inherited boosting learning methods is proposed based on “heredity plus variation” inheriting pattern, which can train four sorts of cascade classifiers. Besides the two traditional cascade classifiers, namely basic cascade classifiers based on “no heredity” inheriting pattern and chained cascade classifiers based on “full heredity” inheriting pattern, there are two new ones, which are feature inherited cascade classifiers and weak classifiers inherited cascade classifiers both based on “partly heredity plus partly variation” inheriting pattern. Although the new ones both have some extra costs, they have better fitting, can balance properly between the convergent speed and the generalization ability and thus outperform the traditional ones. Experimental results on upright frontal face detection based on Real AdaBoost, Gentle AdaBoost and LUT weak classifiers confirm the effectiveness of the new inherited boosting learning methods.