How to Exploit the Transferability of Learned Image Compression to Conventional Codecs

Lossy image compression is often limited by the simplicity of the chosen loss\nmeasure. Recent research suggests that generative adversarial networks have the\nability to overcome this limitation and serve as a multi-modal loss, especially\nfor textures. Together with learned image compression, these two techniques can\nbe used to great effect when relaxing the commonly employed tight measures of\ndistortion. However, convolutional neural network based algorithms have a large\ncomputational footprint. Ideally, an existing conventional codec should stay in\nplace, which would ensure faster adoption and adhering to a balanced\ncomputational envelope.\n As a possible avenue to this goal, in this work, we propose and investigate\nhow learned image coding can be used as a surrogate to optimize an image for\nencoding. The image is altered by a learned filter to optimise for a different\nperformance measure or a particular task. Extending this idea with a generative\nadversarial network, we show how entire textures are replaced by ones that are\nless costly to encode but preserve sense of detail.\n Our approach can remodel a conventional codec to adjust for the MS-SSIM\ndistortion with over 20% rate improvement without any decoding overhead. On\ntask-aware image compression, we perform favourably against a similar but\ncodec-specific approach.\n