Turbo principles meet compression: Rethinking nonlinear transformations in learned image compression

Learned image compression (LIC) has emerged as a powerful approach for achieving high rate–distortion performance. Most existing LIC techniques attempt to address performance limitations associated with downsampling and quantization-induced information loss by employing intricate nonlinear transformations and increasing the feature dimensions in entropy models. In this paper, we introduce a novel perspective by modeling the quantizer as a generalized channel with uniform noise, shifting LIC design toward minimizing the channel’s negative impact on compact feature representations. Drawing inspiration from turbo codes, we propose a turbo-like nonlinear transformation (TLNT). On the encoder side, TLNT-E disperses information loss through parallel component coding units, random interleaving, and puncturing, preserving the integrity of encoded features. At the decoder side, TLNT-D iteratively refines feature representations through interactive processing, enabling accurate reconstruction. Experimental results show that our method outperforms several state-of-the-art nonlinear transformation techniques while maintaining efficiency in parameter count and computational complexity. • A new LIC view formulates quantization as a generalized channel, informing nonlinear transform design. • A turbo-like nonlinear transform (TLNT-E) enhances encoder robustness with parallel units, interleaving, and puncturing. • A decoder-side turbo-like nonlinear transformation (TLNT-D) facilitates collaborative decoding and improved reconstruction.