Extremely low bitrate image compression with decoder-side enhancement for human and machine

To support both machine vision tasks and human visual reconstruction, various image compression frameworks have been proposed. However, due to the growing complexity of models and the limited resources of edge devices, Decoder-side Enhancement Methods (DEM) have been proposed. As the existing DEM fails to simultaneously achieve low bitrates, high task performance, and high reconstruction quality, we propose a decoder-side enhanced image compression framework for extremely low bitrates, where the decoded latent features are separately enhanced for human perception and machine tasks. Thus, our method can support both human and machine tasks with a single bitstream. For classification, we design a Feature Enhancement Transformation Module, which enhances the representational capacity of quantized latent features and better aligns them with the input of the downstream task network, thereby improving classification performance. For reconstruction, a Latent Feature Enhancement Module and a Gated Feature Fusion Module are proposed to recover high-frequency details and reduce over-smoothing and blur. In addition, adversarial training is employed to further improve perceptual quality. Experimental results show that the proposed method outperforms both traditional and end-to-end image compression methods in terms of classification accuracy and human perceptual quality under extremely low bitrates.