改进的多尺度Retinex算法与人脸检测应用

Improved multi-scale Retinex algorithm and its application in face detection

针对多尺度Retinex算法计算速度慢且放大较暗区域噪声的缺陷,提出了改进的多尺度Retinex算法并用于人脸检测。为减少估计照度图像时迭代计算复杂度,根据图像清晰度确定快速导向滤波参数并用于计算照度图像;为抑制较暗区域噪声,根据图像的亮度信息计算各尺度亮度自适应加权系数;为减少照度图像过度平滑引入的估计误差,使用快速导向滤波对反射图像进行平滑。仿真结果表明:改进的算法具有实时性,相较原算法清晰度提高了3.14%,增强后的图像检测率提高了6%。

An improved multi-scale Retinex algorithm is proposed and used in face detection,aiming at problem of slow computing speed and amplifies noises in dark area of multi-scale Retinex. To reduce computing iteration calculation,illuminance image is calculated by fast guided filtering,whose parameters is set by the clarity of image. To suppress the noises in dark area,adaptive weighting factor of each scale is calculated according to the brightness information; To eliminate the estimated error,reflection image is smoothed by fast guided filtering. The simulation results show that the proposed algorithm is real-time. The clarity and detection rate is improved by 3. 14 % and 6 %.