摘要
Viola和Jones提出的基于Boosted Cascade人脸检测方法具有速度快的优点,能够满足实时检测的要求.然而,由于Vio-la的方法采用的是完整的正面人脸作为训练集,所以在检测时对人脸的对称性要求很高.如果这种对称性因为光照、遮挡、旋转的原因而被破坏时,原方法的准确性就会降低.本文针对这一问题提出基于半边脸训练集的改进方案(即只用人脸的左半边或者右半边脸作为训练集),并分析了其可行性.实验结果表明,该方案在一定程度上解决了上述问题.
The Boosted Cascaded method for fast face detection was proposed by Viola and Jones to meet the real-time requirement in detection speed. However, for this method requires whole frontal faces as training set, the learned detection system depends too heavily on the symmetric property of faces. The performance will be degraded greatly ff the faces in target picture lack symmetry by the reasons of illumination, occlusion, or rotation. To solve these problems, this paper suggests an improvement solution that adopts only half faces （ left or right） as training set, and analyzes its feasibility. Experimental results also show that this solution can work out the three problems above to a certain extent.
出处
《小型微型计算机系统》
CSCD
北大核心
2009年第11期2277-2281,共5页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(60503025)资助
