LAC: A Lightweight Linear-Attention Framework for Efficient Learned Image Compression
Fei Yu, Xinwen Hu · IEEE Access · 2025
Efficient learned image compression is critical for modern imaging applications, which demand a balance between high fidelity, low latency, and computational cost, especially on resource-constrained devices. To address these challenges, we introduce LAC, a Lightweight Attention-based Compression framework. Our contributions are threefold: (1) a novel Parallel Attention–Convolution (PAC) module that efficiently captures both local and global features using a lightweight convolutional branch and a linear-attention branch; (2) a Channel-Gated (CG) unit that reduces entropy coding redundancy; and (3) a Shared Head Matrix (SHM) mechanism that decreases parameters in the attention heads. Extensive experiments show that LAC achieves rate-distortion performance competitive with state-of-the-art (SOTA) methods while being significantly more efficient, reducing the parameter count by 40% and improving inference speed by 13.8% compared to a leading SOTA model. By deliberately prioritizing computational efficiency and low latency over peak rate-distortion scores, LAC presents a practical and effective solution for high-quality, efficient image compression, making it a practical solution for low-latency media and mobile applications where computational resources are limited.