AdaBoost人脸检测定点型优化算法

Fixed-Point Optimization Algorithm of AdaBoost Face Detection

提出一种AdaBoost人脸检测的定点型优化算法,该算法以AdaBoost人脸检测原型算法为基础,分析了Cascade瀑布式级联分类器中弱分类器与强分类器分类计算的特点,有效分解了弱分类器与强分类器的计算过程,从而现实了强分类器与弱分类器相关模型参数有效分离标定。优化算法进一步利用图像积分图及弱分类器计算特点,完成对弱分类器计算过程及相关模型参数的定点型转化;同时,利用强分类器浮点的计算精度要求,完成强分类器计算过程及相关模型参数的定点型转化。该定点型AdaBoost人脸检测方法计算精度逼近原浮点型算法计算精度,保持了较高的人脸检测正确率,并利于后期的SIMD并行计算方法优化,同时,也利于算法在定点型嵌入式设备上的移植与优化。

A new fixed-point optimized algorithm for AdaBoost face detection is proposed. Based on the Ada Boost face detection prototype algorithm, the characteristics of classification calculation of the weak classifiers and strong classifiers in waterfall cascade classifier is analyzed, the computing process of the weak classifiers and the strong classifiers is effectively decomposed, and the effective separation and calibration of the model parameters of the strong classifiers and the weak classifiers are realized. By using integral image and the calculation characteristics of the weak classifier and according to the accuracy requirements of the floating point calculation of strong classifiers, the proposed algorithm realizes the classifier calculation and the transformation of related model parameters. The Ada Boost algorithm has the calculation accuracy approximate to that of the original floating-point algorithm and therefore maintains the higher accuracy of face detection, which will be beneficial for the optimization of SIMD parallel computing method and the transplantation and optimization of the algorithm in the fixed point type of embedded equipments.