Learned Image Compression Framework with Quad-Prior Entropy Model

Feifeng Wang, Yao Zhu, Zhaoyi Tian, Liquan Shen

Abstract: Learned image compression has emerged as a promising alternative to traditional codec standards, achieving superior rate-distortion performance by leveraging deep neural networks. However, balancing coding efficiency and compression performance remains a critical challenge in entropy modeling. While autoregressive models capture rich context, they suffer from slow sequential decoding. Parallel models improve speed but underutilize spatial and channel dependencies. To address this trade-off, we propose a learned image compression framework with quad-prior entropy model based on a quadtree-inspired partitioning strategy. Our method divides the latent representation into four groups along the channel dimension and partitions each group into non-overlapping 2$\times$2 spatial patterns. Entropy coding proceeds in four sequential steps, where each step encodes one position within the block using progressively accumulated context from previous steps. This design enables the model to utilize up to 8 spatial neighbors on average—twice that of prior parallel models—and exploits cross-channel correlations through inter-group context sharing. Moreover, all positions within each step are encoded in parallel, ensuring high computational efficiency. When integrated into an end-to-end compression framework with a main autoencoder network and quantization parameter (QP) embedding for variable bitrate control, the proposed method achieves state-of-the-art performance on benchmark datasets.

Team Name: SHUCodec

Submission Number: 7