Gender Classification using the Gaze Distributions of Observers on Privacy-Protected Training Images

We combine observers' gaze distributions with training images whose head regions have been masked or pixelated. Using the regions that attract human attention improves gender classification even when privacy protection removes information from the images.

When photographs of people are used to train a computer, their head regions may be covered, pixelated, or blurred to protect privacy. These changes remove visual information, so a model trained directly on the processed images may become less accurate at gender classification. We investigate whether people’s viewing patterns can help address this problem.

People can sometimes judge gender from the information that remains after the head region has been altered. We therefore measure where observers look in pedestrian images. We use the resulting gaze distribution in the processing performed before computer-based gender classification, helping the method make use of regions that attract human attention.

We tested training images with three types of head-region processing: masking, pixelation, and blur. The gaze-based method improved gender-classification accuracy under these conditions. Rather than restoring the removed information, it changes how the information still present in the image is used.

• Michiko Inoue, Masashi Nishiyama, Yoshio Iwai,
  Gender Classification using the Gaze Distributions of Observers on Privacy-Protected Training Images, Proceedings of 15th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP), Vol. 4: VISAPP, pp. 149 - 156, February 2020.
• 井上 路子, 西山 正志, 岩井 儀雄,
  プライバシーが保護された訓練画像に対する視線位置分布を用いた性別認識,
  精密工学会誌, vol.85 , no.12, pp. 1094 - 1101, December 2019.