双重Haar-like特征瞳孔检测方法

Pupil Detection Method Based on Double Haar-like Features

为了降低光照和镜面反射对检测算法的影响程度,文中提出一种基于结构特征和边缘特征的瞳孔检测算法。利用双重Haar-like特征检测器进行瞳孔粗定位,针对瞳孔区域中存在眼睫毛等影响瞳孔边缘检测的问题,使用形态像素模型进行边缘去噪,提出瞳孔边缘最长假设以确定边缘像素。在LPW瞳孔数据集上的实验结果表明,双重Haar-like特征在提升粗定位精度的基础上,结合形态学像素模型和边缘假设进行边缘去噪,采用易受干扰的最小二乘椭圆拟合算法,提升了复杂环境下眼部图片的瞳孔检测效果。与Swirski和ExCuSe算法相比,文中算法在精确度上与ExCuSe共同取得最佳,整体检测性能比上述2种算法分别高6%和14%。

In order to reduce the influence of illumination and specular reflection on a detection algorithm,the paper presents the pupil detection algorithm based on structure features and edge features.Firstly,a double Haar-like feature detector is used for rough localization of pupils.Secondly,the morphological pixel manipulation pattern is employed for edge denoising to eliminate the interference around pupils(e.g,eyelashes,etc).Finally,the assumption that the longest edge serves as the contour of a pupil is proposed to determine the edge pixels.The evaluation of the Labelled Pupils in the Wild(LPW)datasets show that with the double Haar-like feature detector which improves the accuracy of rough detection and the morphological pixel manipulation patterns and edge assumption for edge denoising,the pupil detection effect in a complex environment is improved even by the direct least squares fitting ellipse method which is prone to interference.This new algorithm is as good as ExCuSe in accuracy,and its overall performance is 6%and 14%higher than that of the Swirski and ExCuSe algorithms respectively.