Effects of Deep Generative AutoEncoder Based Image Compression on Face Attribute Recognition: A Comprehensive Study

Face Attribute Recognition (FAR) is a computer vision task that has attracted a lot of attention for applications ranging from security and surveillance to healthcare. In real-world scenarios, setting up a FAR system requires an important step, which is image compression because of computational, storage, and transmission constraints. However, severe face image compression not adapted to FAR tasks can affect the accuracy of these latter. In this paper, we investigate the impact of image compression based on deep generative models on face attribute recognition performance. In particular, we present a case study on smile and gender detection by face attribute classification of compressed images. For this purpose, we use QRes-VAE (Quantized ResNet Variational AutoEncoder) for image compression, which is, to the best of our knowledge, the most powerful and efficient VAE model for lossy image compression. Unlike prior studies, we quantify the impact of using deep generative autoencoder for image compression on FAR performance. We also study the impact of varying compression rates on the FAR performance. Results obtained from experiments on the CelebA dataset highlight the potential trade-off between image compression by deep generative autoencoder and FAR performance.