Task-Aware Image Compression: The Rate-Distortion-Perception-Classification Tradeoff

Image compression algorithms are typically optimised for recovering the original image bit-by-bit, that is, to minimise image distortion. This applies both to classical schemes and to modern deep neural network-based algorithms, in which the encoder and decoder are optimised jointly. Images, however, are frequently used for specific downstream tasks, e.g., classification, segmentation, or multimedia. Knowledge of such downstream tasks may be leveraged by the encoder to select better features and achieve better compression ratios.
In this work, we consider a joint source-channel coding scenario in which images are compressed to achieve low values of three different metrics: distortion, perception, and classification accuracy. We prove the existence of a tradeoff between these metrics as well as the channel rate (a proxy for the compression ratio), the Rate-Distortion-Perception-Classification (RDPC) tradeoff. We then propose two image compression algorithms to achieve that tradeoff: the RDPCO algorithm which, under simple assumptions, directly solves the optimisation problem characterising the tradeoff, and an algorithm based on an inverse-domain generative adversarial network (ID-GAN), which is more general and achieves extreme compression. Simulation results corroborate the theoretical findings, showing that both algorithms exhibit the RDPC tradeoff.