Region of interest scalable image compression using semantic communications

Growing consumer demand for media content over a wide range of devices has made scalable image compression vital in today’s media landscape. Image compression is conventionally achieved by means of statistical signal processing, but since recently, deep learning techniques are seen to be widely as well. Capabilities of such systems also enable accurate identification of regions of interest in images, leading optimised performance in most applications. This paper proposes a region-of-interest scalable image compression system using semantic communications, where an autoencoder-based semantic encoder performs the base level compression, while a Semantic Mask Extracting Transformer (SeMExT) enables identification of regions of interest to create enhancement layers with different quality levels using a scalable JPEG encoder. When benchmarked against scalable JPEG across a variety of images, the proposed system demonstrates significantly improved compressive performance. The base layer achieved 61.4 times more compression on average, along with better rate-distortion performance at any given quality level.