摘要
为提升设备端的人脸检测效率,提出一种基于YOLO剪枝的设备端人脸检测方法。首先对YOLO网络进行改进,实现人脸检测与面部关键点回归;然后设计一种模型剪枝策略对网络模型进行剪枝,以降低模型复杂度;最后综合多任务损失函数及训练策略以增强模型的鲁棒性。实验结果表明,通过在人脸检测数据集Widerface上与现有的轻量级人脸检测算法进行比较,基于YOLO剪枝的人脸检测方法不仅提升了人脸检测精度,降低了模型复杂度,而且能很好地满足设备端的需求。
To improve the efficiency of device-side face detection,a device-side face detection method based on YOLO pruning is proposed.Firstly,improve the YOLO network to realize face detection and face key point regression.Then,the model pruning strategy is used to prune the network model to reduce the model complexity.Finally,the robustness of the model is enhanced by integrating the multi-task loss function and training strategy.The experimental results show that,compared with the existing lightweight face detection algorithms on the face detection data set Widerface,the proposed method can improve the detection accuracy and reduce the model complexity.Meanwhile,it can also meet the face detection of the equipment demand.
作者
陈泽
朱范炳
CHEN Ze;ZHU Fan-bing(School of Big Data and Artificial Intelligence,Xinyang University,Xinyang 464000,China)
出处
《软件导刊》
2023年第3期196-200,共5页
Software Guide
基金
信阳学院校级科研项目(2021-XJLYB-010)。
关键词
人脸检测
关键点回归
模型剪枝策略
综合损失函数
face detection
key point regression
model pruning strategy
comprehensive loss function