FaceBoxes: 实时高精度CPU人脸检测器

需积分: 12 184 浏览量更新于2024-09-06 收藏 1.41MB PDF 举报

FaceBoxes: A CPU Real-time Face Detector with High Accuracy 是一篇关注于解决实时人脸检测领域挑战的论文。随着人脸识别技术的显著进步，研究人员致力于在保持高性能的同时，实现在中央处理器（CPU）上的实时性能。传统的高效人脸检测模型往往对计算资源有较高的需求，这成为了一个尚未解决的问题。作者Shifeng Zhang、Xiangyu Zhu等人提出了FaceBoxes，一个专注于在CPU上实现高精度和实时性的人脸检测器。FaceBoxes的设计理念在于轻量级但强大的网络结构，包括Rapidly Digested Convolutional Layers (RDCL) 和 Multiple Scale Convolutional Layers (MSCL)。 RDCL是FaceBoxes的关键创新，它旨在通过设计优化的卷积层来确保在CPU上达到实时运行速度。这种设计允许模型在处理速度和计算效率之间找到一个平衡，即使在资源受限的环境里也能保证快速响应。RDCL的设计可能涉及深度可分离卷积、小核尺寸或更高效的计算策略，以减少计算负担。 MSCL则是另一个重要的组成部分，通过在不同层级上引入多尺度的卷积层，它扩展了模型的接收场域（receptive field），增强了对人脸特征的捕捉能力。这种方法使得FaceBoxes能够适应不同大小和比例的人脸，提高了检测的准确性，特别是对于小人脸和侧脸的检测，这是传统方法中的常见难点。论文的核心贡献在于提出了一种新型的轻量化网络架构，能够在保持高精度的同时，显著提升在CPU上的实时性能。这对于广泛应用在实时监控、智能设备和移动平台的人脸识别系统来说具有重要意义。FaceBoxes的实现展示了在不牺牲准确度的前提下，如何通过巧妙的网络设计和优化来降低计算复杂性，这对于推动人脸检测技术的发展具有积极的影响。

FaceBoxes: A CPU Real-time Face Detector with High Accuracy

Shifeng Zhang Xiangyu Zhu Zhen Lei

Hailin Shi Xiaobo Wang Stan Z. Li

CBSR & NLPR, Institute of Automation, Chinese Academy of Sciences, Beijing, China

University of Chinese Academy of Sciences, Beijing, China

{shifeng.zhang,xiangyu.zhu,zlei,hailin.shi,xiaobo.wang,szli}@nlpr.ia.ac.cn

Abstract

Although tremendous strides have been made in face de-

tection, one of the remaining open challenges is to achieve

real-time speed on the CPU as well as maintain high perfor-

mance, since effective models for face detection tend to be

computationally prohibitive. To address this challenge, we

propose a novel face detector, named FaceBoxes, with supe-

rior performance on both speed and accuracy. Speciﬁcally,

our method has a lightweight yet powerful network struc-

ture that consists of the Rapidly Digested Convolutional

Layers (RDCL) and the Multiple Scale Convolutional Lay-

ers (MSCL). The RDCL is designed to enable FaceBoxes

to achieve real-time speed on the CPU. The MSCL aims at

enriching the receptive ﬁelds and discretizing anchors over

different layers to handle faces of various scales. Besides,

we propose a new anchor densiﬁcation strategy to make

different types of anchors have the same density on the

image, which signiﬁcantly improves the recall rate of small

faces. As a consequence, the proposed detector runs at 20

FPS on a single CPU core and 125 FPS using a GPU for

VGA-resolution images. Moreover, the speed of FaceBoxes

is invariant to the number of faces. We comprehensively

evaluate this method and present state-of-the-art detection

performance on several face detection benchmark datasets,

including the AFW, PASCAL face, and FDDB.

1. Introduction

Face detection is one of the fundamental problems in

computer vision and pattern recognition. It plays an im-

portant role in many subsequent face-related applications,

such as face alignment [46], face recognition [47] and face

tracking [12]. With the great progress over the past few

decades, especially the breakthrough of convolutional neu-

ral network, face detection has been successfully applied in

our daily life under various scenarios.

However, there are still some tough challenges in un-

controlled face detection problem, especially for the CPU

Corresponding author

devices. The challenges mainly come from two require-

ments for face detectors: 1) The large visual variation of

faces in the cluttered backgrounds requires face detectors to

accurately address a complicated face and non-face classi-

ﬁcation problem; 2) The large search space of possible face

positions and face sizes further imposes a time efﬁciency

requirement. These two requirements are conﬂicting, since

high-accuracy face detectors tend to be computationally

expensive. Therefore, it is one of the remaining open issues

for practical face detectors on the CPU devices to achieve

real-time speed as well as maintain high performance.

In order to meet these two conﬂicting requirements, face

detection has been intensely studied mainly in two ways.

The early way is based on hand-craft features. Follow-

ing the pioneering work of Viola-Jones face detector [37],

most of the early works focus on designing robust fea-

tures and training effective classiﬁers. Besides the cascade

structure, the deformable part model (DPM) is introduced

into face detection tasks and achieves remarkable perfor-

mance. However, these methods highly depend on non-

robust hand-craft features and optimize each component

separately, making the face detection pipeline sub-optimal.

In brief, they are efﬁcient on the CPU but not accurate

enough against the large visual variation of faces.

The other way is based on the convolutional neural net-

work (CNN) which has achieved remarkable successes in

recent years, ranging from image classiﬁcation to object

detection. Recently, CNN has been successfully introduced

into the face detection task as feature extractor in the tra-

ditional face detection framewrok [23, 41, 42]. Moreover,

some face detectors [4, 45] have inherited valid techniques

from the generic object detection methods, such as Faster

R-CNN [29]. These CNN based face detection methods

are robust to the large variation of facial appearances and

demonstrate state-of-the-art performance. But they are too

time-consuming to achieve real-time speed, especially on

the CPU devices.

These two ways have their own advantages. The for-

mer has fast speed while the latter owns high accuracy.

To perform well on both speed and accuracy, one natural

下载后可阅读完整内容，剩余8页未读，立即下载

计算机视觉-Archer

粉丝: 2617
资源: 9

FaceBoxes: 实时高精度CPU人脸检测器

我国提出实时、高准确率的CPU面部检测方法——FaceBoxes.pdf

基于深度置信网络的人脸识别方法研究.pdf

基于树莓派与深度学习的人脸识别考勤系统.pdf

人工智能在多媒体设备巡检中的应用研究.pdf

基于改进型FaceBoxes的人头检测算法.docx

FaceBoxes-TF

Python-使用pytorch实现了FaceBoxes

Python-FaceBoxes具有高精度的CPU实时人脸检测器

Python-Faceboxes具有高精度的CPU实时人脸检测器

cpp-FaceBoxes具有高精度的CPU实时人脸检测器

最新资源