基于沃尔什特征和增强型Cascade算法的人脸检测

Fast Face Detection Based on Walsh Feature and Enhanced Cascade Algorithm

在训练基于AdaBoost和Cascade算法的人脸检测器时,由于使用了大量的Haar-Like特征,所以训练过程消耗了大量的时间和存储空间,为此该文提出用较少的沃尔什特征来代替大量的Haar-Like特征,可以较大幅度地降低特征之间的冗余,节约训练时间和存储空间。针对Nesting Cascade完全继承前层分类器的不足之处,提出一种具有自主和继承双重特性的增强型Cascade算法。MIT-CBCL库上的实验表明:沃尔什特征可以加快训练速度,而增强型Cascade算法有助于提高测试精度。最后,使用训练好的人脸检测器对MIT+CMU前视人脸测试集进行了测试,结果证明该文方法比相应的对比方法更加有效。

The training time cost is very expensive when mass Haar-Like features are used to obtain the face detector based on AdaBoost and Cascade algorithm. The paper presents a Walsh feature to replace Haar-Like feature in the training process, which can decrease the redundancy among the features and save the training time and memory. Aiming at the shortage of entirely inheriting prior classifters in the Nesting Cascade algorithm, an enhanced Cascade algorithm that has independence characteristic and inheritance speciality is proposed. The experimental results from MIT-CBCL database show that the Walsh feature can accelerate the training process and the enhanced Cascade algorithm can increase the test precision. A trained face detector is used to detect faces in the MIT ＋ CMU test set, and the detected results demonstrate that the proposed algorithm is more effective than other correlative methods.