Despite a short history, neural image codecs have been shown to surpass classical image codecs in terms of rate-distortion performance. However, most of them suffer from significantly longer decoding times, which hinders the practical applications of neural image codecs. This issue is especially pronounced when employing an effective yet time-consuming autoregressive context model since it would increase entropy decoding time by orders of magnitude. In this paper, unlike most previous works that pursue optimal RD performance while temporally overlooking the coding complexity, we make a systematical investigation on the rate-distortion-complexity (RDC) optimization in neural image compression. By quantifying the decoding complexity as a factor in the optimization goal, we are now able to precisely control the RDC trade-off and then demonstrate how the rate-distortion performance of neural image codecs could adapt to various complexity demands. Going beyond the investigation of RDC optimization, a variable-complexity neural codec is designed to leverage the spatial dependencies adaptively according to industrial demands, which supports fine-grained complexity adjustment by balancing the RDC tradeoff. By implementing this scheme in a powerful base model, we demonstrate the feasibility and flexibility of RDC optimization for neural image codecs.

翻译：尽管历史短暂，神经图像编解码器在率失真性能上已被证明超越传统图像编解码器。然而，大多数神经图像编解码器面临解码时间显著更长的问题，这阻碍了其实际应用。当采用有效但耗时的自回归上下文模型时，这一问题尤为突出，因为该模型会使熵解码时间增加几个数量级。与以往多数研究在暂时忽略编码复杂度的前提下追求最优率失真性能不同，本文系统研究了神经图像压缩中的率-失真-复杂度（RDC）优化。通过将解码复杂度量化为优化目标中的一个因子，我们能够精确控制RDC权衡，进而展示神经图像编解码器的率失真性能如何适应各种复杂度需求。在RDC优化研究的基础上，我们设计了一种可变复杂度神经编解码器，能够根据工业需求自适应地利用空间依赖性，通过平衡RDC权衡实现细粒度的复杂度调整。通过将该方案部署于强大的基础模型，我们证明了神经图像编解码器中RDC优化的可行性与灵活性。