【作者】 王旭；

【导师】 李冬松；

【作者基本信息】 哈尔滨工业大学 ， 计算数学， 2020， 硕士

【摘要】 随着现代人工智能信息技术的不断飞速发展,人脸识别等新技术技术逐渐得到兴起,人脸检测技术也变得愈发重要。早期的各种人脸图像检测技术主要用于研究的对象是具有很强约束力和条件的各种人脸检测图像,例如那些无任何背景、无遮挡的图像,因此这些人脸很容易的被定位和找到。由于人工智能的不断发展。人脸检测与识别技术等已成为研究的重点。但是由于有些场景,例如在枪击摄像头拍摄的监控场景中。人脸目标距离摄像设备较远或由于光照、遮挡等其他原因。由此所致而面临的一些列问题也使得目前人脸图像检测技术开始作为独立的研究课题一并受到许多研究者的重视。人脸检测技术的难点有很多。主要有以下三个方面。人脸可以有不同角度的遮挡。成像的角度不同,导致图像侧脸或脸部旋转等。人脸有不同的肤色。成像的环境条件不同。如光照条件、有无阴影等。因此本文对于低分辨率和有遮挡的人脸研究进行了如下工作:首先,阐述了MTCNN网络的框架,对它的三阶段级联的模块做了介绍。提出了基于MTCNN的网络结构。对输入到P-Net网络的图片进行预处理。接着对P-Net、R-Net、O-Net网络进行了改进。并且此方法在WIDER FACE、FDDB数据集上取得了很好的效果。接下来使用语义分割网络来辅助人脸挖掘与检测。可以有效提升困难人脸的检出率语义分割网络进行研究,语义分割网络用于目标检测,这里将语义分割网络用于人脸检测。当人脸有不同程度遮挡时。可以辅助人脸检测。挖掘人脸周围的信息。使用Fast-SCNN网络,它采用一个学习下采样模块,一个标准全局特征提取器处理模块,一个全局特征数据融合处理模块,以及一个标准的分类器。所有的模块都是用深度可分离卷积构建。可以有效地提高困难人脸的检出率。最后提出了用残差网络连接SSD网络的人脸关键点五点定位的方法。网络的前置模块使用的是残差模块。它的网络性能远超传统网络模型。后面连接一个SSD网络。SSD网络是也是使用CNN进行检测。它使用了多种数据增强的方法,包括水平翻转、剪裁、放大、缩小等。SSD提取出人脸框。接着对人脸关键点进行五点定位。得到精准的人脸框。并且可以看到,在COFW数据集上的效果较好。因此对有遮挡人脸的检测的优越性。综上所述,本文是对低分辨率和有遮挡的人脸进行研究。但是这些方法都存在或多或少的局限性。但是本文的方法可以处理大多数人脸。并且可以进行同时检测和定位。更多还原

【Abstract】 With the rapid development of modern artificial intelligence information technology,new technologies such as face recognition have gradually emerged,and face detection technology has become increasingly important.The early various face image detection technologies were mainly used for researching various face detection images with strong binding and conditions,such as those without any background and no occlusion,so these faces are easily Locate and find.Due to the continuous development of artificial intelligence.Face detection and recognition technology has become the focus of research.However,due to some scenes,such as surveillance scenes shot by shooting cameras.The face target is far away from the camera device or due to light,occlusion and other reasons.A series of problems faced by this also make the current face image detection technology begin to be paid attention by many researchers as an independent research topic.There are many difficulties in face detection technology.There are three main aspects.Faces can be blocked at different angles.The angle of imaging is different,which leads to the side of the image or the rotation of the face.Human faces have different skin tones.The environmental conditions for imaging are different.Such as lighting conditions,the presence or absence of shadows,etc.Therefore,this paper has carried out the following work on the low-resolution and occlusion face research:First,the framework of the MTCNN network is described,and its three-stage cascaded modules are introduced.A network structure based on MTCNN is proposed.Pre-process the pictures input to the P-Net network.Then the P-Net,RNet and O-Net networks were improved.And this method has achieved good results on WIDER FACE and FDDB data sets.Next,use semantic segmentation network to assist face mining and detection.The semantic segmentation network that can effectively improve the detection rate of difficult faces is studied.The semantic segmentation network is used for target detection.Here,the semantic segmentation network is used for face detection.When the face is blocked in varying degrees.Can assist face detection.Mining the information around the face.Using Fast-SCNN network,it uses a learning downsampling module,a standard global feature extractor processing module,a global feature data fusion processing module,and a standard classifier.All modules are constructed with deep separable convolutions.It can effectively improve the detection rate of difficult faces.Finally,a five-point positioning method for the key points of the face connected with the SSD network by the residual network is proposed.The premodule of the network uses a residual module.Its network performance far exceeds the traditional network model.Connect an SSD network behind.SSD networks are also detected using CNN.It uses a variety of data enhancement methods,including horizontal flip,crop,zoom in,zoom out,etc.The SSD extracts the face frame.Then five-point positioning of key points on the face.Get a precise face frame.And you can see that the effect on the COFW dataset is better.Therefore,it is superior to the detection of blocked faces.To sum up,this article is to study the face with low resolution and occlusion.These methods have more or less limitations.But the method in this article can deal with most faces.And can be detected and located at the same time.更多还原