EntroMamba:Efficient Entropy Modeling for Learned Image Compression via Selective State Spaces

Entropy modeling, which predicts the probability distribution of the quantized latent representation, is pivotal for rate-distortion optimization in learned image compression. However, existing CNN- and Transformer-based entropy approaches face a critical trade-off: capturing long-range dependencies incurs prohibitive computational overhead, while efficient local models sacrifice global context. To resolve this dilemma, we proposed a Mamba-based entropy approach, termed EntroMamba, which jointly optimizes modeling capacity and inference efficiency through two innovations: (1) HyperMamba2D, a hyperprior extractor to capture long-range spatial dependencies via 2D selective state-space scanning; (2) HyConMaskMamba, a causal dual-branch module that fuses local (via masked convolution) and global autoregressive context (via masked Mamba). Experimental results show that the proposed method achieves superior rate–distortion performance over state-of-the-art learned image compression approaches while maintaining a favorable balance between compression efficiency and computational complexity.