face detection人脸检测技术简介

合集下载
  1. 1、下载文档前请自行甄别文档内容的完整性,平台不提供额外的编辑、内容补充、找答案等附加服务。
  2. 2、"仅部分预览"的文档,不可在线预览部分如存在完整性等问题,可反馈申请退款(可完整预览的文档不适用该条件!)。
  3. 3、如文档侵犯您的权益,请联系客服反馈,我们会尽快为您处理(人工客服工作时间:9:00-18:30)。

Loss function
MTCNN: cross-eeuclidean metricntropy loss Faceboxes: 2-class softmax loss for classification Smooth L1 loss for regression
future prospects
2.Method and Architecture :
MTCNN
Faceboxes
3.Anchor densification strategy:
4.Train and Experiment:
Training dataset
WIDER FACE FDDB CelebA
Data augmentation
1. Goal of the works :
Faceboxes
CPU only Train and practise without GPU
Real Time Implement real-time by RDCL-Rapidly Digested Convolutional Layers
High Accuracy Implement highaccuracy by MSCLMultiple Scale Convolutional Layers
NMS
Non-maximum suppression——to remove the redundant boxes which point to the same face
Bbox-Regression
To approach the true boxes by zooming and moving
Example
基于人脸检测的课堂分析
Face Detection
17-18学年度第二次汇报 唐康
Main Process
02
Box sliding
04
Bounding box (regression)
01
Figure Input
03
Detection(classification)
1 Figure Input
Color distortion Random cropping Scale transformation Horizontal flipping Face-box filter
Matching strategy
Match each face to the anchor with the best jaccard overlap; Match anchors to any face with jaccard overlap higher than a threshold
Development History of Face Detection:
Priori knowledge and Feature Based
e.g.: Methods based handcraft features—shape of faces , color of skins , etc
Machine learning
e.g.: SVM based NN based Adaboost based— “Viola-Jones” method
Deep learning
e.g.:
R-CNN (series) YOLO (v1~v5) SSD
IoU
Intersection-over-Union
Data set
FDDB Wider Face FaceDB MTFL Data Augmentation Color distortion Random cropping Scale transformation Horizontal flipping Face-box filter …
2 Box Sliding
2.FaceBoxes
A CPU Real-time Face Detector with High Accuracy Authors: Shifeng Zhang, Xiangyu Zhu, Zhen Lei, Hailin Shi, Xiaobo Wang, Stan Z. Li Accepted by:IJCB 2017
Face Recognition
Mood Recognition
Signing System
Class Evaluation System
THANKS
Parameters while box sliding
Size of the boxes
Stride Min and max of the face boxes Image pyramid
Something else Convolution YOLO strategy …
3 Detection Algorithm
1.MTCNN
Multi-task Cascaded Convolutional Networks
Authors: Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao Accepted by: IEEE Signal Processing Letters, 2016, 23(10):1499-1503.
相关文档
最新文档