LIC_TCM-Light: Lightweight Design and Research for TCM Image Compression Technology

In image compression, the Transformer-CNN Hybrid (TCM) module in the Learned Image Compression (LIC) framework balances local and non-local feature modeling, but existing lightweight TCM models still have inference efficiency bottlenecks in mobile/real-time scenarios, failing to meet real-time needs. This study optimizes the TCM module’s lightweight design while keeping the original framework’s overall architecture unchanged, improving feature extraction, hyper-prior modeling, and entropy coding via modular decomposition and hybrid mechanism. It solves high computational complexity of traditional Transformer-based models without significant compression performance loss—dynamic block partitioning enables lightweight high-dimensional feature entropy modeling, and the hybrid module realizes efficient local-global feature interaction. Experiments show the proposed model has 16.68 M parameters (62.9% less than small TCM), 53.23 GMac FLOPs (74.85% less), and 1.7× faster inference. On datasets like Kodak, PSNR loss is <0.7 dB at the same bit rate, balancing efficiency and performance, offering a solution for edge device image compression and a reference for Transformer-based model lightweight design.