Successive learned image compression: Comprehensive analysis of instability

Recently, learned image compression relying on deep learning has shown remarkable improvement compared to conventional image compression. While the existing studies on learned image compression focus only on single-step compression, this paper addresses the problem of successive image compression (SIC), which refers to repeated application of encoding and decoding an image. It can be commonly involved in the processes of image sharing, editing, and re-distribution, but the performance of learned image compression during SIC has not been studied in literature. Thus, we provide comprehensive analysis of successive learned image compression in three aspects. First, we newly introduce the issue of instability of learned compression methods during SIC, and analyze its causes and affecting factors. Second, we provide SIC benchmarking studies for the state-of-the-art learned compression methods, based on which we compare different methods and investigate the components affecting the instability in detail. Finally, we propose two methods to mitigate the issue of instability of learned methods during SIC.