人脸分析算法审美观的规训与偏向:基于计算机视觉技术的智能价值观实证研究

Discipline and Bias in the Aesthetics of Face Analysis Algorithms:An Empirical Research on AI Values Based on Computer Vision Technology

在人工智能正越来越多地辅助人类进行决策的背景下,算法的价值判断行为值得学术界关注。本研究聚焦于人脸分析算法的审美价值判断行为,选取智能化人脸分析应用Face++作为研究对象,使用1247张在控制环境下拍摄的人脸正面视图对Face++中的“颜值评估”及“人脸美化”算法进行逆向测试。研究使用计算机视觉技术计算目标人脸的视觉年龄、瘦削程度、肤色白皙程度等面部属性,计算其与算法“颜值评估”结果的关联,并比较“人脸美化”算法是否带来上述属性的显著改变。结果证实“颜值评估”算法推崇“瘦、白、幼”的片面审美观,而“人脸美化”算法将这种单一的审美偏好付诸自动化的美学实践。在上述发现的基础上,研究进一步探讨了算法审美价值背后的意识形态偏向、算法与人类审美实践相交互的运作逻辑。

When artificial intelligence is more widely used for assisting humans in making decisions,the value judgment behavior of algorithms deserves the attention of the academic community. This study focused on the esthetic preference of the face analysis algorithms. We chose intelligent face analysis application Face++ and used 1,247 frontal photos of faces taken in a controlled environment to reverse-test the “beauty evaluation” and “face beautify” algorithms provided by Face++. Research found that the algorithm value judgment inherited social value preference and discipline human beings. The study used computer vision technology to calculate the visual age, thinness, and whiteness of faces, and explored their correlation with the automatic “beauty evaluation” score, and examined whether the “face beautify” algorithm results in significant changes to the above attributes. The results confirmed that the “beauty evaluation” algorithm promotes the aesthetics of “thin, white, and young”, and the “face beautify” algorithm further puts this aesthetic preference into an automated aesthetic practice, strengthening the existing preference. Based on the results above, the research further explores the ideological preference behind the aesthetic value of algorithms, and the operational logic of the interaction between algorithms and human aesthetic practices.