人脸图像分析与妆容图像合成

CASIA OpenIR > 毕业生 > 硕士学位论文

	人脸图像分析与妆容图像合成
	赵昱程
	2018-05
学位类型	工学硕士
中文摘要	个人肖像图像是社交网络图片分享的重要组成部分，然而受拍摄环境、摄影设备和用户自身摄影技巧等因素的影响，实景拍摄的人像图片质量普遍不高。图像合成技术被广泛应用于生成符合用户个人特点且具美学吸引力的高质量图像。作为妆容合成美化的基础，人脸图像分析技术通过对人像进行建模完成信息抽取及解析。现有基于人脸关键点的妆容合成方法，以人像五官关键点为基础进行妆容虚拟合成，在处理面部皮肤及头发等区域时普遍存在信息残缺等问题，妆容合成质量及稳定性有待提高。本文提出基于人脸图像语义分割的妆容合成。考虑到人脸结构属性，人脸图像分析相关任务如人脸检测、关键点定位和语义分割之间存在一定协同性，本文基于深度多任务学习框架对人脸图像进行分析研究，并在此基础上展开基于人脸图像语义分割结果的妆容合成方法研究。本文的主要研究工作和贡献如下：首次提出一种基于深度卷积神经网络的人脸关键点定位和语义分割多任务联合分析框架。在多任务框架中引入语义分割增强模块，以得到更精细的分割结果；提出基于残差连接的语义分割增强模块，增进网络不同分支信息共享。研究深度多任务学习框架下的人脸检测和语义分割问题。提出多尺度残差连接特征融合模块，提升梯度传递效率，在不增加参数的情况下增加模块非线性拟合能力；研究多任务网络中降采样特征抽取模块对模型性能的影响。构建人脸妆容合成应用。提出基于人脸关键点定位和语义分割结果的人脸妆容合成方法，基于人脸检测和语义分割结果，对皮肤平滑和染发应用进行研究，基于人脸关键点定位结果，完成换脸应用。
英文摘要	Portrait images play an important role in information sharing in social networks. However, due to restricted factors such as shooting environment, photography equipment or users' photography skills, the quality of portrait shots is generally not satisfying. Image synthesis techniques are widely used to generate high quality images that are more aesthetically appealing to user's personal preference. Face image analysis techniques, as the basis of makeup, perform facial information extraction and analysis by statistically modeling the human portraits. Existing face alignment based visual methods of makeup synthesis perform poorly when dealing with some facial regions such as facial skin and hair. The quality and stability of these methods need to be improved. In this thesis, we propose a novel makeup synthesis method based on semantic face segmentation. Taking the attributes of facial structure and related tasks of face analysis methods such as face detection, face alignment and face segmentation into account, we propose to use deep multi-task learning framework. To further validate our model, we show applications of makeup synthesis. The main contributions are summarized as follows: To the best of our knowledge, we are the first to present a deep multi-task framework that solves face alignment and face segmentation tasks jointly. We incorporate refinement module into our multi-task learning framework in order to parse the contours of facial parts more exactly. With a carefully designed refinement residual module, the cross-layer features are fused in a collaborative manner. We present a deep multi-task learning framework for face detection and face segmentation. A multi-scale residual module with good gradient flow is proposed to make the decision function more discriminative without excessive increase in the parameters of the network. We study the effect of downsampling module on the performance of the multi-task network. We design applications based on the face analysis results. Virtual makeup is achieved based on the result of face alignment and face segmentation. A skin smoothing and hair dyeing application is performed based on face detection and face segmentation result. Then the face swap can be completed based on face alignment result.
关键词	多任务学习人脸检测关键点定位语义分割图像合成
学科领域	计算机视觉
文献类型	学位论文
条目标识符	http://ir.ia.ac.cn/handle/173211/20983
专题	毕业生_硕士学位论文
作者单位	中国科学院自动化研究所
第一作者单位	中国科学院自动化研究所
推荐引用方式 GB/T 7714	赵昱程. 人脸图像分析与妆容图像合成[D]. 北京. 中国科学院大学,2018.

条目包含的文件
文件名称/大小	文献类型	版本类型	开放类型	使用许可
人脸图像分析与妆容图像合成-签名.pdf（10395KB）	学位论文		限制开放	CC BY-NC-SA