Face Detection, Alignment, Quality Assessment and Attribute Analysis with Multi?Task Hybrid Convolutional Neural Networks
2019-12-30 09:44:36 《ZTE Communications》 2019年3期
GUO Da ZHENG Qingfang PENG Xiaojiang LIU Ming
Abstract： This paper proposes a universal framework， termed as Multi?Task Hybrid Convolutional Neural Network （MHCNN）， for joint face detection， facial landmark detection， facial quality， and facial attribute analysis. MHCNN consists of a high?accuracy single stage detector （SSD） and an efficient tiny convolutional neural network （T?CNN） for joint face detection refinement， alignment and attribute analysis. Though the SSD face detectors achieve promising results， we find that applying a tiny CNN on detections further boosts the detected face scores and bounding boxes. By multi?task training， our T?CNN aims to provide five facial landmarks， facial quality scores， and facial attributes like wearing sunglasses and wearing masks. Since there is no public facial quality data and facial attribute data as we need， we contribute two datasets， namely FaceQ and FaceA， which are collected from the Internet. Experiments show that our MHCNN achieves face detection performance comparable to the state of the art in face detection data set and benchmark （FDDB）， and gets reasonable results on AFLW， FaceQ and FaceA.
Keywords： face detection; face alignment; facial attribute; CNN; multi?task training
http：//kns.cnki.net/kcms/detail/34.1294.TN.20190920.2104.004.html， published online September 20， 2019
Manuscript received： 2019?06?11
ace analysis has been widely?used in many applications such as face beautification system， face based access system， and video anti?terrorism system. Although great progresses have been made in recent， detecting and aligning abnormal faces such as occlusion faces as well as analyzing their attributes in surveillance are very challenging due to low resolution， lack of abnormal training data， etc.
Generally， face detection， face alignment， facial quality assessment， and facial attribute recognition are considered as separate face analysis tasks which may have their own task?dependent models. For face detection， from traditional Viola?Jones （VJ） face detector  and deformable part model （DPM） based face detector  to recent convolutional neural networks （CNN） based face detectors -， the performance of face detection has been improved significantly. Among all the CNN based face detectors， those detectors evolved from anchor?based object detectors （e.g. single shot multibox detector （SSD） ， Faster Region CNN （R?CNN） ）， such as single shot scale?invariant face detector （S3FD）  and Face Region CNN （R?CNN） ， are superior to pure CNN face detectors ， because anchor?based detectors can naturally leverage the context information. For face alignment， CNN based methods have also achieved promising results -. However， most of the alignment methods must be initialized by the provided face bounding box in advance， which presents a great demand of joint face and landmark detection ， . For facial quality assessment， traditional methods  mainly apply local binary patterns （LBP）  or histograms of oriented gradients （HOG）  features with a support vector machine （SVM） classifier， while a few works with CNN obtain state?of?the?art performance ， . For facial attribute recognition，  introduces the CelebA dataset with 40 facial attributes ranging from smiling to gender， and subsequently many deep learning based methods are developed for facial attribute analysis -. Unfortunately， CelebA does not contain the attribute of wearing face mask which we are interested in.
In this paper， we address several face analysis tasks including face detection， face alignment， facial quality assessment， and facial attribute recognition in the wild. Specifically， we propose a Multi?Task Hybrid Convolutional Neural Network （MHCNN） which unifies all the tasks in a framework. MHCNN is comprised of two parts， namely a single stage detector （SSD） and an efficient tiny CNN （T?CNN）. Compared to pure CNN face detectors， the SSD based face detector ensures high baseline accuracy on challenging face images in the wild. Instead of performing multi?task learning with SSD like ， we apply a tiny CNN which is more feasible for multi?task face analysis. We argue that a CNN operated on cropped faces brings complementary information to SSD. Given an image， the MHCNN first detects all the faces with a SSD based face detector and then refines both the scores and bounding boxes with T?CNN. Since the T?CNN is applied in individual faces， it is straightforward to add multiple tasks upon it. We here add face alignment， facial quality assessment， and facial attribute recognition. In addition， we introduce a facial attribute dataset， i.e. FaceA， which contains two highly?concerned attributes in surveillance， namely wearing sunglasses and wearing masks. We also introduce a human?based facial quality assessment dataset， i.e. FaceQ， where low?quality cases include occlusion， low?resolution， large pose， etc. We evaluate our face detection performance on the well?known face detection data set and benchmark （FDDB） dataset， and demonstrate our T?CNN on our FaceA and FaceQ.
The remained of this paper is organized as follows. In Section 2， we review related work on face detection and multi?task learning. We introduce our MHCNN and its training strategy in Section 3. Our collected datasets are introduced in Section 4. We present experimental results in Section 5 and conclude the paper in Section 6.
2 Related work
We mainly review the face detection and multi?task learning for face analysis in this section. One can refer to - and  for face image quality assessment and facial attribute recognition， respectively.
2.1 Face Detection
Face detection has been a well?studied field of computer vision. According to the used features， face detection methods can be roughly divided into two categories， namely hand?craft feature based methods and CNN feature based methods.
1） Hand?craft feature based methods. The cascaded face detector proposed by VIOLA et al.  （VJ detector） obtains good performance in simple scenarios with real?time efficiency. Due to the relatively weakness of Haar?like features， the VJ detector degrades significantly in real?world applications with larger visual variations of human faces. Some works improved the VJ detectors by replacing the Haar?like features with more advanced hand?crafted ones -， which need more computational cost. Another popular pipeline of face detection is based on DPM ， ， . It performs relatively better than VJ detector in the wild but it is more computationally expensive and usually requires expensive annotation in the training stage.
2） CNN feature based methods. Since the remarkable success of CNN in object classification ， many progresses have been made for face detection -. These CNN?based methods can be mainly concluded as three categories， namely cascaded CNN based face detection， two?stage region?based face detection， and single?stage face detection. The cascaded CNN based face detection pipeline， which inherits the advantage of the VJ detector， utilizes several small networks from simple to complex to detect faces and regress face boxes in a coarse?to?fine manner -. Two?stage region?based face detection pipeline is mainly transferred from region?based object detectors， like R?CNN ， Fast R?CNN ， and Faster R?CNN . This method mainly includes two stages， namely proposal generation and classification. The single?stage face detection pipeline directly generates face boxes and scores from dense anchor boxes ， . The face detection model for finding tiny faces  trains separate detectors for different scales. S3FD  presents multiple strategies to improve the performance of small faces. Single stage headless （SSH）  models the context information by large filters on each prediction module. PyramidBox  utilizes contextual information with improved SSD network structure. The advantage of single?stage face detectors is that it can use the context semantic information to assist in detecting faces， which is difficult for cascade face detectors. So， we introduce single?stage detection architecture into cascade face detector to get higher performance.
2.2 Multi?Task Learning
There are some existing works attempting to jointly solve the problem of face detection， alignment and facial attribute in a single model. YANG et al.  train deep convolution neural networks for facial attribute recognition to obtain high response in face regions which further yield candidate windows of faces. ZHANG et al.  proposed to use facial attribute recognition as an auxiliary task to enhance face alignment performance using deep convolutional neural network. CHEN  et al. apply random forest based on the features of pixel value difference to jointly conduct alignment and detection， but these handcraft features are low?level features and greatly limit its performance. Multitask cascaded convolutional network （MTCNN）  leverages a cascaded architecture with three stages of shallow to deep convolution networks to jointly predict face and landmark locations in a coarse?to?fine manner， but the performance of MTCNN is limited by cascade architecture. So， we propose MHCNN with single?stage architecture in cascade face detector and joint multi?task learning to improve the performance of face detector.
4） Scale compensation anchor matching. To match more tiny faces for anchors， S3FD decreases the jaccard overlap threshold from 0.5 to 0.35 in order to increase the average number of matched anchors， and further sorts these anchors with jaccard overlap higher than 0.1 and selects top?N as matched anchors.
5） Training. We use the training set of the WIDER FACE  to train the SSD?based detector， and use the same data augmentation strategies as S3FD， including color distort， random crop， and horizontal flip. The input size of network is fixed to 640×640. We use smooth L1 loss for face bounding box regression and softmax loss for face/non?face classification. We apply non?maximum suppression （NMS） to remove the highly overlapped results with a threshold of 0.7.
3.3 The Multi?Task Tiny CNN
We design an efficient T?CNN for the second part of MHCNN. The T?CNN aims to further refine the candidates from the SSD?based face detector， detect facial landmarks， assess the face quality， and recognize two importance facial attributes.
Fig. 3 presents the architecture of T?CNN. This architecture is inspired by the MTCNN . Specifically， we use off?the?shelf O?Net of MTCNN as the architecture while add more tasks. T?CNN takes as input the 48×48 face regions， and output results for four face tasks as follows.
1） Face classification. We find that there are a number of non?face cases in the detections of the first part of MHCNN， which are mainly caused by hard negative contexts and low qualities， such as a person with back face. These cases can be relaxed by directly classifying the face regions. We believe that adding a refinement T?CNN could be complementary with the SSD?based face detector. The face scores from both S3FD and T?CNN will be averaged to provide the final detections.
2） Landmark localization. We also predict five landmarks at eyes， nose， and mouth as in MTCNN. As shown in MTCNN， adding landmark localization is helpful for face recognition. We explain that landmarks can be viewed as a post validation of faces.
3） Attribute classification. Facial attribute is naturally a multi?label task. In this paper， we only concern about two important attributes for surveillance applications， i.e. wearing face masks and wearing sunglasses. A sigmoid layer is responded to each attribute.
4） Face quality classification. Face quality can impact the face/non?face classification scores in practice. We add face quality task as a two?class （i.e. high quality and low quality） classification problem since it is hard to annotate accurate quality scores for human. We consider two issues for face quality classification： a） face quality assessment served as a filter for subsequent face recognition system since too many low?quality and unrecognizable faces can impact the communication of front devices and cloud devices; b） As shown in Fig. 3， we use a separate fully?connected （FC） layer for face quality assessment because this task mainly depends on non?semantic information and it brings negative influence to other tasks if sharing the same FC layer in practice.
5） Training. Since T?CNN is partially served as a face/non?face refinement of the SSD?based detector in our MHCNN framework， we need to collect training data according to the results of SSD?based detector. To this end， we first calculate the Intersection?over?Union （IoU） ratio between the detections from the SSD?based detector and ground?truth faces on the training set of WIDER FACE， and then select these detections with IoU above 0.4 as positives and those less than 0.35 as negatives. The number of total face/non?face training data for T?CNN is 60000 with a ratio of 1：3. For facial landmark localization， we use the CelabA dataset which annotated with five facial landmarks， and apply random crop and gaussian blur as two data augmentation strategies. Euclidean loss is used for training facial landmark regression. We use our FaceA and FaceQ to train facial attribute recognition and face quality assessment. As for the training of facial attributes， the sigmoid layer with binary cross?entropy loss is used.
4 FaceA and FaceQ
This section details the collection and annotation of our FaceA and FaceQ datasets. To our knowledge， there is no public dataset for face attribute recognition with both wearing face masks and wearing sunglasses， and there is also no public dataset for face quality assessment in the wild. To meet our research， we collect the FaceA and the FaceQ datasets， and will make it publicly available to promote this area.
1） Collection. We make a python script and start from crawling “wearing sunglasses” and “wearing face mask” in image searching engine such as baidu （www.baidu.com） and google （www.google.com）. We find it is hard to collect a large scale of data for wearing face mask since this case usually happens in surveillance. Totally， we crawled 1 409 and 1 335 images for “wearing sunglasses” and “wearing face mask”， respectively. After crawling， we then feed these images into the first part of our MHCNN and crop the detected faces for further annotation.
2） FaceA. With the detected faces， we find there are many noises which are not human faces （e.g. cartoon and animation） or without the expected attributes. We manually remove these noises， and finally the FaceA dataset consists of 1 072 faces with sunglasses and 663 faces with face mask. The FaceA also includes a background category which contains 630 faces without wearing things. We randomly split both classes with 8：2 as training set and test set. Fig. 4 shows some examples of FaceA. We note that these faces are mostly with large head poses which could be challenging for recognition.
3） FaceQ. With the crawled images， we find there are a lot of faces that neither belong to “wearing sunglasses” nor “wearing face mask”， and that there are a number of faces with either high or low resolution. Thus， we collect FaceQ from the same source with FaceA but with three rules：
· Face resolution： Blurred and tiny faces are divided into low?quality class.
· Head pose： We collect faces as high?quality class only if their eyes can be seen clearly and their resolution are larger than 80×80. In other word， profile faces are not selected as high?quality class.
· Occlusion： Occluded faces are selected as the low?quality class except for the faces that are only occluded by wearing sunglasses and masks.
After manually selection， we totally obtain 1001 high?quality faces and 1 097 low?quality faces. We also randomly split both classes with 8：2 as training set and test set. Some examples of FaceQ are shown in Fig. 5.
In this section we first present the implementation details， and then analyze the effectiveness of our joint multi?task training for the face detection， and further evaluate the final model on FDDB face detection benchmark and our own benchmarks. Finally， we evaluate the time of inference of our MHCNN.
5.1 Implementation Details
We use Caffe toolbox for implementation of our MHCNN. For the SSD?based face detector， we follow the training setting of S3FD， using the pretrained VGG16 to initialize， and the other layers are randomly initialized with the “Xavier” method . We fine?tune the pretrained model using stochastic gradient descent （SGD） with 0.9 momentum， 0.0005 weight decay and batch size 32. We train 80k iterations by using 10-3 learning rate， then continue training for 20k iterations with 10-4 and 10-5 learning rates. For T?CNN， we convert different datasets to hdf5 format for joint multi?task training， and train the model using SGD with 0.9 momentum and 0.0005 weight decay. The batch size of each task is 512， and we concatenate all data for joint training. Due to the fact that face quality assessment task depends on different information （i.e. low?level information） compared to the other tasks， we train T?CNN in two stages. First， we train the face classification， facial landmark localization， and face attribute recognition tasks with 10-1 learning rate for 200 iterations. Then we freeze the weights and only train the face quality recognition part of T?CNN with 10-1 learning rate for another 500 iterations.
5.2 Evaluation on Face Detection Task
We evaluate and compare the face detection performance of our MHCNN on the FDDB dataset. FDDB contains 5 171 face annotations in 2 845 images. We compare our face detector to the state?of?the?art methods ， ， ， ， -. Table 1 shows the recall ratio at 2 000 false positive and Fig. 6 compares the receiver operating characteristic （ROC） curves to several state?of?the?art methods. Although our T?CNN aims for multi?task facial analysis， compared to the original S3FD， our extra T?CNN provides complementary information for face/non?face classification which boosts S3FD by around 0.3%. It is worth noting that a tiny improvement is difficult on the nearly?saturated FDDB. From Fig. 6， we observe that our MHCNN mainly improve the true positive rate at low false positive rate， which means the MHCNN has higher scores for true faces than S3FD. This character is practical in real applications.
5.3 Ablation Study of T?CNN
We make an ablation study of our T?CNN on the facial attribute task. We perform experiments on the collected FaceA dataset， and use the sigmoid scores for each attribute with the best threshold searched on the training set.
Table 2 presents the results of ablation study on FaceA. We find several observations from Table 2. First， adding the facial landmark localization task improves both attribute tasks， where the gains for sunglasses and face mask are 0.43% and 2.6%， respectively. Second， training with all attribute tasks and landmark localization achieves the best results on FaceA. Third， the overall results with multiple tasks are relatively high though we only use a low?resolution input， which demonstrates the efficiency of our T?CNN.
1） Visualization on FaceA. Fig. 7 visualizes some false positives on FaceA. We find that “wearing mask” is easily confused by large?pose faces with heavy hair， partial occlusion， and sunglasses; “wearing sunglasses” is confused by wearing common glasses， partial occlusion， and wearing mask.
2） Face quality assessment with T?CNN. For face quality assessment task， we evaluation our T?CNN on the FaceQ dataset. We compare a well?known and popular method in real applications， i.e. LBP feature with SVM. In this method we use the circular LBP operator with 8 sampling points in a circular area with radius 2， and divide face image into 7×5 to calculate LBP histogram， getting a 2 065 dim feature vector for each face image. Using the LBP face image features， we train a SVM model with radial basis kernel function （RBF） to predict either the normalized comparison scores. The cost of SVM is set at 1.5 and the gamma for RBF at 6.82. Table 3 presents the comparison between MHCNN and LBP+SVM. We find that T?CNN outperforms LBP+SVM by 3.34%， which demonstrates its effectiveness. As a traditional method， LBP+SVM is also promising on this task which may be explained by that the face quality assessment task mainly depend on low?level texture information.
3） Visualization on FaceQ. Fig. 8 shows some false positives on FaceQ. We find that most of the faces with serious occlusion by sunglasses and masks are recognized as low?quality faces， which may make sense since they are not suitable for recognition by human; several smooth profile faces also have low quality scores since they have little texture information; these low?quality faces with small occlusion by sunglasses are easily categorized into high?quality faces， which may be explained by the fact that these faces can provide relatively rich texture information.
5.4 Inference Time
During inference， we set 0.5 as the face confidence threshold in both parts of MHCNN. We perform the inference in 10 real?world surveillance images with 1 080×1 920 scales and report the average time. We first downscale the images to 320×568 and then use our MHCNN. The inference time of the SSD?based detector and T?CNN are around 22 ms/frame and 2 ms/frame in NVIDIA TITAN Xp， respectively. Overall， our MHCNN can run 40 FPS in NVIDIA TITAN Xp for four face tasks including resizing computational time.
In this paper， we propose MHCNN for face detection， facial landmark detection， facial quality， and facial attribute analysis. We combine the single stage detector and CNN?based detector to boost the performance of face detection and implement multi?task learning. Our MHCNN achieves real?time performance in NVIDIA GPU for four face tasks. Additionally， we contribute two datasets on face attribute and face quality assessment. Experiments show that our MHCNN achieves the state?of?the?art on FDDB benchmark and gets reasonable results on FaceQ and FaceA.
 VIOLA P， JONES M J. Robust Real?Time Face Detection [J]. International Journal of Computer Vision， 2004， 57（2）： 137-154. DOI： 10.1023/B：VISI.0000013087.49260.fb
 MATHIAS M， BENENSON R， PEDERSOLI M， et al. Face Detection Without Bells and Whistles [C]//European Conference on Computer Vision. Zurich， Switzerland， 2014： 720-735. DOI： 10.1007/978?3?319?10593?2_47
 ZHU C， ZHENG Y， LUU K， et al. CMS?RCNN： Contextual Multi?Scale Region?Based CNN for Unconstrained Face Detection [M]. Deep Learning for Biometrics. Cham， Switzerland： Springer， 2017： 57-79
 JIANG H， LEARNED?MILLER E. Face Detection with the Faster R?CNN [C]//12th IEEE International Conference on Automatic Face & Gesture Recognition （FG 2017）. Washington DC， USA， 2017： 650-657. DOI： 10.1109/FG.2017.82
 LI H， LIN Z， SHEN X， et al. A Convolutional Neural Network Cascade for Face Detection [C]//IEEE Conference on Computer Vision and Pattern Recognition. Boston， USA， 2015： 5325-5334. DOI： 10.1109/CVPR.2015.7299170
 YANG S， LUO P， LOY C C， et al. From Facial Parts Responses to Face Detection： A Deep Learning Approach [C]//IEEE International Conference on Computer Vision. Santiago， Chile， 2015： 3676-3684. DOI： 10.1109/ICCV.2015.419
 ZHANG K， ZHANG Z， LI Z， et al. Joint Face Detection and Alignment Using Multitask Cascaded Convolutional Networks [J]. IEEE Signal Processing Letters， 2016， 23（10）： 1499-1503. DOI： 10.1109/LSP.2016.2603342
 ZHANG S， ZHU X， LEI Z， et al. Faceboxes： A CPU Real?Time Face Detector with High Accuracy [C]//2017 IEEE International Joint Conference on Biometrics （IJCB）. Denver， Colorado， USA， 2017： 1-9. DOI： 10.1109/BTAS.2017.8272675
 NAJIBI M， SAMANGOUEI P， Chellappa R， et al. SSH： Single Stage Headless Face Detector [C]//IEEE International Conference on Computer Vision. Venice， Italy， 2017： 4875-4884. DOI： 10.1109/ICCV.2017.522
 LIU W， ANGUELOV D， ERHAN D， et al. SSD： Single Shot Multibox Detector [C]//European Conference on Computer Vision. Amsterdam， The Netherlands， 2016： 21-37. DOI： 10.1007/978?3?319?46448?0_2
 REN S， HE K， GIRSHICK R， et al. Faster R?CNN： Towards Real?Time Object Detection with Region Proposal Networks [C]//Advances in Neural Information Processing Systems. Montreal， Canada， 2015： 91-99. DOI：10.1109/TPAMI.2016.2577031
 ZHANG S， ZHU X， LEI Z， et al. S3FD： Single Shot Scale?Invariant Face Detector [C]//IEEE International Conference on Computer Vision. Venice， Italy， 2017： 192-201. DOI： 10.1109/ICCV.2017.30
 GLOROT X， BENGIO Y. Understanding the Difficulty of Training Deep Feedforward Neural Networks [C]//13th International Conference on Artificial Intelligence and Statistics. Sardinia， Italy， 2010： 249-256.
 SUN Y， WANG X， TANG X. Deep Convolutional Network Cascade for Facial Point Detection [C]//IEEE Conference on Computer Vision and Pattern Recognition. Portland， USA， 2013： 3476-3483. DOI： 10.1109/CVPR.2013.446
 ZHU X， LEI Z， LIU X， et al. Face Alignment Across Large Poses： A 3D Solution [C]//IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas， USA， 2016： 146-155. DOI：10.1109/CVPR.2016.23
 FENG Z H， KITTLER J， AWAIS M， et al. Wing Loss for Robust Facial Landmark Localisation with Convolutional Neural Networks [C]//IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City， USA， 2018： 2235-2245. DOI： 10.1109/CVPR.2018.00238
 ZHUANG C， ZHANG S， LEI Z， et al. FLDet： A CPU Real?Time Joint Face and Landmark Detector [C]// IAPR International Conference on Biometrics （ICB）. Crete， Greece， 2019
 BHARADWAJ S， VATSA M， SINGH R. Can Holistic Representations be Used for Face Biometric Quality Assessment？ [C]//IEEE International Conference on Image Processing. Melbourne， Australia， 2013： 2792-2796. DOI： 10.1109/ICIP.2013.6738575
 OJALA T， PIETIK?INEN M， M?ENP?? T. Multiresolution Gray?Scale and Rotation Invariant Texture Classification with Local Binary Patterns [J]. IEEE Transactions on Pattern Analysis & Machine Intelligence， 2002 （7）： 971-987. DOI： 10.1109/TPAMI.2002.1017623
 DALAL N， TRIGGS B. Histograms of Oriented Gradients for Human Detection [C]//International Conference on Computer Vision & Pattern Recognition （CVPR'05）. San Diego， USA， 2005， 1： 886-893. DOI： 10.1109/CVPR.2005.177
 HERNANDEZ?ORTEGA J， GALBALLY J， FIERREZ J， et al. FaceQnet： Quality Assessment for Face Recognition Based on Deep Learning [DB/OL]. （2019?04?03）. https：//arxiv.org/abs/1904.01740
 NASROLLAHI K， MOESLUND T B. Face Quality Assessment System in Video Sequences [C]//European Workshop on Biometrics and Identity Management. Roskilde， Denmark， 2008： 10-18. DOI： 10.1007/978?3?540?89991?4_2
 LIU Z， LUO P， WANG X， et al. Deep Learning Face Attributes in the Wild [C]//IEEE International Conference on Computer Vision. Santiago， Chile， 2015： 3730-3738. DOI： 10.1109/ICCV.2015.425
 HAN H， JAIN A K， WANG F， et al. Heterogeneous Face Attribute Estimation： A Deep Multi?Task Learning Approach [J]. IEEE Transactions on Pattern Analysis and Machine Intelligence， 2018， 40（11）： 2597-2609. DOI： 10.1109/TPAMI.2017.2738004
 RANJAN R， SANKARANARAYANAN S， CASTILLO C D， et al. An All?in?One Convolutional Neural Network for Face Analysis [C]//12th IEEE International Conference on Automatic Face & Gesture Recognition （FG 2017）. Washington DC， USA， 2017： 17-24. DOI： 10.1109/FG.2017.137
 ZHANG Z， LUO P， LOY C C， et al. Learning Deep Representation for Face Alignment with Auxiliary Attributes [J]. IEEE Transactions on Pattern Analysis and Machine Intelligence， 2015， 38（5）： 918-930. DOI： 10.1109/TPAMI.2015.2469286
 BEST?ROWDEN L， JAIN A K. Learning Face Image Quality from Human Assessments [J]. IEEE Transactions on Information Forensics and Security， 2018， 13（12）： 3064-3077. DOI： 10.1109/TIFS.2018.2799585
 ZHANG L， CHU R， XIANG S， et al. Face Detection Based on Multi?Block LBP Representation [C]//International Conference on Biometrics. Seoul， South Korea， 2007： 11-18. DOI： 10.1007/978?3?540?74549?5_2
 ZHU Q， YEH M C， CHENG K T， et al. Fast Human Detection Using a Cascade of Histograms of Oriented Gradients [C]//IEEE Computer Society Conference on Computer Vision and Pattern Recognition （CVPR06）. New York， USA， 2006， 2： 1491-1498. DOI： 10.1109/CVPR.2006.119
 PHAM M T， GAO Y， HOANG V D D， et al. Fast Polygonal Integration and its Application in Extending Haar?Like Features to Improve Object Detection [C]//IEEE Computer Society Conference on Computer Vision and Pattern Recognition. San Francisco， USA， 2010： 942-949. DOI： 10.1109/CVPR.2010.5540117
 YAN J， LEI Z， WEN L， et al. The Fastest Deformable Part Model for Object Detection [C]//IEEE Conference on Computer Vision and Pattern Recognition. Columbus， USA， 2014： 2497-2504. DOI： 10.1109/CVPR.2014.320
 RAMANAN D， ZHU X. Face Detection， Pose Estimation， and Landmark Localization in the Wild [C]//IEEE Conference on Computer Vision and Pattern Recognition （CVPR）. Rhode Island， USA， 2012： 2879-2886. DOI： 10.1109/cvpr.2012.6248014
 KRIZHEVSKY A， SUTSKEVER I， HINTON G E. Imagenet Classification with Deep Convolutional Neural Networks [C]//Advances in Neural Information Processing Systems. Lake Tahoe， USA， 2012： 1097-1105. DOI： 10.1145/3065386
 GIRSHICK R， DONAHUE J， DARRELL T， et al. Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation [C]//IEEE Conference on Computer Vision and Pattern Recognition. Columbus， USA， 2014： 580-587. DOI： 10.1109/CVPR.2014.81
 GIRSHICK R. Fast R?CNN [C]//IEEE International Conference on Computer Vision. Santiago， Chile， 2015： 1440-1448. DOI： 10.1109/ICCV.2015.169
 TANG X， DU D K， HE Z， et al. Pyramidbox： A Context?Assisted Single Shot Face Detector [C]//European Conference on Computer Vision （ECCV）. Munich， Germany， 2018： 797-813. DOI： 10.1007/978?3?030?01240?3_49
 ZHANG Z， LUO P， LOY C C， et al. Facial Landmark Detection by Deep Multi?Task Learning [C]//European Conference on Computer Vision. Zurich， Switzerland， 2014： 94-108. DOI： 10.1007/978?3?319?10599?4_7
 CHEN D， REN S， WEI Y， et al. Joint Cascade Face Detection and Alignment [C]//European Conference on Computer Vision. Zurich， Switzerland， 2014： 109-122. DOI： 10.1007/978?3?319?10599?4_8
 SIMONYAN K， ZISSERMAN A. Very Deep Convolutional Networks for Large?Scale Image Recognition [DB/OL]. （2014?09?04）. https：//arxiv.org/abs/1409.1556
 YANG S， LUO P， LOY C C， et al. Wider Face： A Face Detection Benchmark [C]//IEEE Conference on Computer Vision and Pattern Recognition. Las Vega， USA， 2016： 5525-5533. DOI： 10.1109/CVPR.2016.596
 YANG B， YAN J， LEI Z， et al. Aggregate Channel Features for Multi?View Face Detection [C]//IEEE International Joint Conference on Biometrics. Clearwater， USA， 2014： 1-8. DOI： 10.1109/BTAS.2014.6996284
 YAN J， ZHANG X， LEI Z， et al. Face Detection by Structural Models [J]. Image and Vision Computing， 2014， 32（10）： 790-799. DOI： 10.1016/j.imavis.2013.12.004
 MARKUS N， FRLJAK M， PANDZIC I S， et al. Object Detection with Pixel Intensity Comparisons Organized in Decision Trees [DB/OL]. （2013?05?20）. https：//arxiv.org/abs/1305.4537
 LI H， LIN Z， BRANDT J， et al. Efficient Boosted Exemplar?Based Face Detection [C]//IEEE Conference on Computer Vision and Pattern Recognition. Columbus， USA， 2014： 1843-1850. DOI： DOI： 10.1109/CVPR.2014.238
 LI J， ZHANG Y. Learning Surf Cascade for Fast and Accurate Object Detection [C]//IEEE Conference on Computer Vision and Pattern Recognition. Portland， USA， 2013： 3468-3475. DOI： 10.1109/CVPR.2013.445
GUO Da received the B.Eng. from the Computer Engineering College， JiMei University， China in 2018. He is currently a master student at the Shenzhen Institutes of Advanced Technology， Chinese Academy of Sciences， China. His research direction is face detection and recognition based on deep learning.
ZHENG Qingfang received the B.S. degree in civil engineering from Shanghai Jiao Tong University， China in 2002 and Ph.D. degree in computer science from Institute of Computing Technology， Chinese Academy of Science， China in 2008. He is currently the chief scientist of video technology with ZTE Corporation. His research interests include computer vision， multimedia retrieval， image/video processing， with a special focus on low power embedded application and large?scale cloud application.
PENG Xiaojiang （[email protected]） received his Ph.D. from School of Information Science and Technology from Southwest Jiaotong University， China in 2014. He currently is an associate professor at the Shenzhen Institutes of Advanced Technology， Chinese Academy of Sciences， China. He was a postdoctoral researcher at Idiap Institute， Switzerland from 2016 to 2017， and was a postdoctoral researcher in LEAR Team， INRIA， France， working with Prof. Cordelia Schmid from 2015 to 2016. He serves as a reviewer for IJCV， TMM， TIP， CVPR， ICCV， AAAI， IJCAI， FG， Image and Vision Computing， IEEE Signal Processing Letter， Neurocomputing， etc. His research focus is in the areas of action recognition and detection， face recognition， facial emotion analysis， and deep learning.
LIU Ming received the M.Sc. degree from Harbin Engineering University， China in 2011. He is currently a senior engineer with ZTE Corporation. His research interests include object detection， tracking and recognition.
- High Speed Polarization?Division Multiplexing Transmissions Based on the Nonlinear Fourier Transform
- RAN Centric Data Collection for New Radio
- A Lightweight Sentiment Analysis Method
- A Service?Based Intelligent Time?Domain and Spectral?Domain Flow Aggregation in IP?over?EON Based on SDON
- Reinforcement Learning from Algorithm Model to Industry Innovation: A Foundation Stone of Future Artificial Intelligence
- Big Data-Driven Residents’ Travel Mode Choice: A Research Overview