改进的多目标回归实时人脸检测算法

针对物体检测实时多目标回归算法中分别优化各四个位置参数，割裂了四个位置变量之间的关系，造成对物体的边框回归不够准确且训练不易收敛的问题，提出一种带检测评价函数（Intersection over Union，IoU）作为损失函数的实时多目标回归人脸检测算法。首先基于Redmond等提出实时多目标回归模型，采用该模型检测实时性的机制，然后融合了IoU函数作为位置参数的损失函数，将实时多目标回归模型中的四个独立位置参数整合成一个单元进行优化，避免了基础模型的缺陷。算法在人脸检测基准库FDDB上进行测试，实验结果表明：在人脸检测的有效性上优于主流的传统人脸检测算法，检测速度上领先于其他经典深度学习方法。提出的算法在检测人脸的有效性和检测速度两者之间取得了一个较好的平衡，为构建实用的人脸相关应用系统提供了参考价值。

Improved multi-objective regressive real-time face detection algorithm

Real-time multi-objective regression algorithm optimizes four location parameters respectively and separates four location variables. It makes localization less accurate and training hard to converge. Based on the problems, this paper proposes a real-time multi-object regression face detection algorithm, combining detection evaluation function Intersection over Union（IoU） as loss function. Based on multi-objective regression algorithm model proposed by Redmond etc., it adopts its real-time detection mechanism. Then IoU is introduced as loss function of location parameter. Four independent location parameters in real-time multi-object regression model are integrated into one unit and optimized, avoiding the defects of the base model. The algorithm is tested in face detection benchmark FDDB. The results indicate that this algorithm is superior to a mainstream traditional one in terms of face detection’s effectiveness, and it outperforms other classical deep learning methods in terms of detection speed. The algorithm achieves a balance between face detection’s effectiveness and detection speed. It provides some reference value to construct practical face applications.