分散矩形特征结合软级联的人眼检测研究

Study on Eye-Detection Combined Scattered Rectangle Feature with Soft Cascade

采用优化的分散矩形特征结合软级联进行了人眼检测研究,针对分散矩形特征会造成特征数量暴增的问题,给出了一种优化解决方案;针对级联分类器训练过程中训练难度不断加大的问题,提出了一种按指数规律调整每层的最小检出率和最大误检率的方法 .优化后训练所需时间缩短为优化前的1/3,改进后分类器性能相比原始特征有所提高.人眼检测实验结果表明:该分类器具有更好的分类能力.

Optimized scattered rectangle feature combining with soft cascade is used in eye detection. First, because the number of features will explode by using scattered rectangle feature, an optimized solution is proposed. Second, for the increasingly difficulty of cascade classifier training process, a method that the minimum hit rate and the maximum false alarm of every stage is adjusted according to exponential curve has been proposed. The improved methods result in two- thirds fewer training time consume and a higher classifier performance. In eye detection experiments, the results show that the classifier has better classification ability.