Diffusion Probabilistic Model Based Face Anonymization in Embedded Environments

Computer vision systems depend on vast and diverse datasets to achieve optimal performance. However, in scenarios where data collection involves personal identifiable information, privacy-protecting legislation can impede the acquisition process. To address this challenge, we present a novel face anonymization pipeline capable of running on embedded platforms. The proposed methodology combines denoising diffusion probabilistic models (DDPMs) and advanced image super-resolution techniques, such as DFDNet and ESPCN, to achieve high-quality anonymized facial images while preserving crucial facial keypoint locations. We conduct a comprehensive evaluation of the pipeline's performance, encompassing image quality, keypoint precision, and efficiency on two popular single-board computers, namely the Nvidia Jetson Nano and the Raspberry Pi 4B. The source code is available at https://github.com/balazsmorv/facediffusion.