非约束环境下基于YOLOv3-Tiny多模型融合方式的快速人脸定位算法  

Fast face location algorithm based on YOLOv3-Tiny multi-model fusion in unconstrained environment

作　　者：万秋波 屈金山 朱泽群 Wan Qiubo;Qu Jinshan;Zhu Zequn(China Three Gorges University,College of Computer and Information Technology,Yichang 443002,China)

机构地区：[1]三峡大学计算机与信息学院,湖北宜昌443002

出　　处：《长江信息通信》2021年第1期75-77,共3页Changjiang Information & Communications

摘　　要：非约束环境下的人脸定位算法是诸多面部感知应用系统中的关键基础模块,一直是极具挑战性的课题。随着海量训练数据集的出现及深度学习技术的发展,基于深层卷积神经网络的视觉目标检测方法取得了突破性的进展,YOLOv3-Tiny是其中一种具有较高准确率的快速通用目标检测算法,但由于其输出神经元的物理感受野范围会随网络输入尺寸的固定而被限定,使其无法在具有尺度范围跨度过大的非约束人脸检测中充分发挥其检测性能。为了有效扩展YOLOv3-Tiny网络检测神经元的有效感知范围,文章提出了一种基于YOLOv3-Tiny多模型融合方式的快速人脸定位算法。首先根据人脸尺寸对原始图像集的标签数据进行筛选,划分为不同的子集,并利用它们分别对多个具有不同有效感知范围的YOLOv3-Tiny模型进行训练。接着,利用多模型对输入图像独立进行推理,并基于非极大值抑制算法及相应的尺度约束阈值实现检测结果的有效融合。实验结果显示,该算法能够有效利用多模型各自的检测优势,实现跨度尺度范围下的无约束人脸检测,具有重要应用潜力。Face localization algorithms in unconstrained environments are the key basic modules in many face-sensing application systems, and have always been a challenging subject. With the emergence of massive training data sets and the development of deep learning technology, the visual target detection method based on deep convolutional neural networks has made breakthrough progress. YOLOv3-Tiny is one of the fast and general detection algorithm with high accuracy, but because the physical receptive field range of its output neuron will be limited with the fixed network input size, it cannot give full play to its detection performance in the non-constrained face detection with excessive scale range. In order to effectively expand the effective perception range of YOLOv3-Tiny network detection neurons, the article proposes a fast face location algorithm based on YOLOv3-Tiny multi-model fusion method. First, filter the label data of the original image set according to the face size, divide it into different subsets, and use them to train multiple YOLOv3-Tiny models with different effective perception ranges. Then, the multi-model is used to infer the input image independently, and the effective fusion of the detection results is realized based on the non-maximum suppression algorithm and the corresponding scale constraint threshold.The experimental results show that the algorithm can effectively use the respective detection advantages of multiple models to achieve unconstrained face detection under the span scale range, which has important application potential.

关 键 词：非约束环境 多模型融合 人脸定位 深度学习 YOLOv3-Tiny 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]