Face De-Identification for Research and Clinical Use

描述(由申请人提供)：这项申请回应了美国国立卫生研究院促进数据共享和患者隐私的呼吁。共享录制的视频的一个主要障碍是需要保护参与者的身份。同样，对耻辱的担忧是许多需要心理健康服务的人(例如在军队中)未能做到这一点的一个原因。我们提出了一种在视频中识别患者和研究参与者身份的系统。人脸去身份识别自动将人脸表情从源人脸图像(保密)转移到目标人脸图像(非机密)。该系统保护了人脸匿名性，同时保留了原始源视频的面部表情。然后，目标视频可以传达源人的情感、交流意图、痛苦以及神经或生理状态，而不显示源人的脸。人脸识别将使研究人员和临床医生能够在不损害隐私或机密性的情况下共享视频档案。此外，该系统的一个版本可能被用来在基于互联网的采访中保护隐私和匿名性。创新。该项目有四项创新。该方法(1)去除了身份信息，同时保留了面部动力学；从而保留了面部的信息价值，以传达情感、疼痛和相关状态。(2)适应微妙和自发的面部动作，而不是只模仿一些预定义的磨牙表情(例如，高兴或悲伤)。(3)不需要目标人员进行培训步骤。(4)不需要对视频进行人工标注。这个系统是完全自动的。接近。该软件将把带有对象(源)的脸的视频作为输入，并自动生成或输出带有去识别的脸的视频。该项目将使用新的机器学习和计算机视觉算法，仅使用目标受试者的一张正面图像，将微妙的面部表情从源受试者(原始视频)转移到目标受试者。这种方法的一个主要新奇之处是使这个过程完全自动化。该算法将使用商用的人脸识别软件和定制的面部表情分析软件进行验证。

DESCRIPTION (provided by applicant): This application addresses NIH's call to promote data sharing and patient privacy. A major obstacle to sharing of recorded video has been the need to protect participants' identity. Similarly, concern about stigma is a reason that many people in need of mental health services (e.g., in the military) fail to do so. We propose a system to de-identify patients and research participants in video. Face de-identification transfers facial expression automatically from source face images, which are confidential, to target face images, which are not. The system safeguards face anonymity while preserving the facial expression of the original source video. The target video then can communicate the emotion, communicative intent, pain, and neurological or physiological status of the source person without displaying the source person's face. Face de-identification would enable video archive sharing among researchers and clinicians without compromising privacy or confidentiality. Moreover, a version of this system could potentially be used to preserve privacy and anonymity in internet-based interviews. Innovation. The project has four innovations. The approach (1) Removes identity information while retaining facial dynamics; thus preserving the information value of the face to communicate emotion, pain, and related states. (2) Accommodates subtle and spontaneous facial actions, rather than imitating only some predefined molar expressions (e.g., happy or sad). (3) Requires no training steps by target persons. (4) And requires no hand annotation of video. The system is entirely automatic. Approach. The software will take as input a video with the face of a subject (source) and automatically generate or output a video with the face de-identified. The project will use new machine learning and computer vision algorithms for transferring subtle facial expression from a source subject (original video) to a target subject, using only one frontal image of the target subject. A major novelty of the approach is to make the process completely automatic. The algorithm will be validated using commercially available software for face recognition and custom software for facial expression analysis.