【作者】 李俊；

【导师】 杨若瑜；

【作者基本信息】 南京大学 ， 计算机科学与技术， 2019， 硕士

【摘要】 人脸检测和人脸关键点定位是计算机视觉领域备受关注的两个任务,它们在现实生活中有着重要的应用。智能手机利用人脸检测来优化人物拍摄效果,人脸关键点定位可以辅助安全驾驶。除此之外,人脸检测和人脸关键点定位在人脸识别和表情识别中扮演重要的角色。正脸无遮挡的人脸检测和定位已经取得了很好的效果,但是在现实环境下,受到姿态、光照、遮挡等因素的影响,人脸呈现不同的表观,检测和定位变得极具挑战性。本文提出了三种方法,提升了无约束环境下人脸检测和人脸关键点定位的准确率。第一,本文借助多任务学习,针对人脸检测和关键点定位任务设计了新的网络结构。不同于之前的方法联合学习人脸/非人脸分类、人脸框回归和关键点定位任务,本文额外引入了判断两张图片是否属于同一类的判别任务(如果两张图片同为人脸或者非人脸,标签为1,否则为0),并把网络改成二输入多输出结构,还为不同任务选择了不同的特征,提高了人脸检测和关键点定位的准确率。该方法在FDDB和WIDER FACE数据集上取得了很好的效果,同时在人脸关键点定位数据集AFLW上也展现了不错的性能。第二,人脸周围环境信息有助于检测小尺度、模糊、部分遮挡以及头部偏转的人脸。因此本文借助语义分割方法来挖掘人脸周围环境信息。在语义分割任务中,每个像素都被标注了类别,借助这种强监督方式,可以有效提升困难人脸的检出率。考虑到语义分割模型速度慢,为了适应人脸检测,本文设计了新的语义分割网络并使用LIP服装解析数据集进行训练。对比其他人脸检测方法在FDDB数据集上的正确率,语义分割模型显示出很大的优越性。第三,本文通过忽略大姿态下不存在的关键点和被遮挡的关键点,实现了能同时处理大姿态和部分遮挡的人脸关键点定位。该方法先借助本文提出的多任务学习得到人脸关键点,再利用轮廓、角点信息提取关键点部件框,最后针对部件分别训练回归模型得到关键点坐标。由于只处理了真正存在的关键点,并引入了部件框的分别定位,因而提高了人脸关键点定位的准确性,且弥补了人脸框形状差异和人脸框偏移对定位精度的不良影响。该方法在AFLW和COFW数据集上超过了以前一些最优方法,同时,在LFPW数据集上准确率也很有竞争力。综上所述,本文首先通过联合学习和新的网络结构,提升人脸检测和关键点定位的准确率,然后再分别利用语义分割和基于人脸部件的回归来进一步解决如遮挡、大姿态等困难情况下的人脸检出和关键点定位问题。实验证明,本文方法能准确并快速处理绝大多数人脸,并同时完成检测和定位,应用范围更广。更多还原

【Abstract】 Face detection and face alignment are well studied problems in computer vision and they are widely used in real life.Smart phones utilize face detection to take better pictures of people.Face alignment can help safe driving.Besides,face detection and face alignment play important role in face recognition and expression recognition.Many face detectors can easily detect near frontal faces without occlusion and under the circumstances,face alignment can accurately locate key points.But in real-word conditions,robust face detection and face alignment are very challenging due to the large variability of face caused by occlusion,pose and illumination.In this paper,we propose three methods and successfully improve the accuracy of face detection and alignment.First,this paper designs new network for face detection and face alignment by multi-task learning.Differing from other methods which jointly learn face classification,bounding box regression and facial landmark localization,our method adds another task to predict weather two pictures are of the same kind(if both two pictures are faces or non-faces,the label is one.Otherwise,the label is zero)and transforms the model form single input to two inputs.It also chooses different features according to assignments and improves the accuracy of face detection and alignment.This method outperforms the state-of-the-art methods on FDDB and WIDER FACE datasets.As for face alignment,it is also superior to other methods on LFPW dataset.Second,parts around face can help to detect small,vague and occluded faces.Therefore,this paper proposes to use semantic segmentation model to make the best of parts around face.In semantic segmentation,every pixel is marked.By strong-supervised learning,we can detect more small and occluded faces.Considering the slow speed of semantic segmentation models,this paper designs a new segmentation model for face detection.The method outperforms other methods on FDDB dataset which is trained on LIP dataset.Thirdly,in this paper,we achieve face alignment under large poses and partial occlusion by ignoring the points which cannot be seen under occlusion and extreme poses.The method first gets facial parts by the outputs of the multi-task learning proposed by this paper and then it further utilizes corner points and contours to get more facial parts.Finally we train regression models for every facial part to get key points.Because of only locating the truly existing points and utilizing facial parts to obtain key points,it leads to higher accuracy and reduces the adverse influence of the drift and shape of face detection results.The method outperforms the state-of-the-art methods on AFLW and COFW datasets and it is also comparable to other methods on LFPW dataset.In conclusion,this paper first utilizes multi-task learning and designs new networks to improve the accuracy of face detection and face alignment.Then,in order to detect faces and locate key points under occlusion and large poses,this paper takes advantages of semantic segmentation and facial parts separately.Experiments validate that the methods proposed by this paper can handle most of faces and accomplish face detection and alignment simultaneously.Therefore,it can be more widely used.更多还原