VersaCodec: A Versatile Codec With Variable Compression Rate for Switchable Machine Perception of Edge Devices

The Image Coding for Machine (ICM) approach effectively mitigates challenges posed by data transmission of machine vision tasks on edge devices with bandwidth-limited networks. However, existing ICM solutions consume tremendous storage in complex and multi-task scenarios, which is impractical for edge devices in the field of Internet of Things (IoT). To address this issue, we introduce VersaCodec, a versatile codec with variable compression rates, built upon the Swin Transformer architecture. VersaCodec utilizes a quality-map-based encoding-decoding scheme and an adapted prompt technique to maximize the preservation of semantic information for machine perception tasks via exploring the spatial importance. In VersaCodec, a quality-map generator is designed to extract the spatial semantic information from conditions and tasks. VersaCodec enables seamless task switching and incorporates a conditional input mechanism, allowing flexible adjustment of the compression rate. This makes VersaCodec adaptive to varying environmental conditions and thus suitable for specific application scenarios of IoT. Experiments demonstrate that our VersaCodec outperforms TransTIC tailored for specific compression rates across multiple downstream tasks.