LLSC: End-to-End Image Semantic Communication Framework for Low-Light Scenarios

Transmitting images captured under low-light conditions suffers from dual degradation: intrinsic quality loss from poor illumination and extrinsic impairments from bandwidth-limited channels. This paper presents a novel end-to-end semantic communication framework explicitly designed for low-light visual recovery, which jointly optimizes semantic feature extraction, channel transmission, and perceptual reconstruction. The proposed system employs Retinex-based decomposition to extract illumination and reflectance as semantic representations, followed by a content-adaptive pyramid encoder that hierarchically captures multi-scale features across four levels with cross-scale interaction modules. A progressive decoder reconstructs enhanced images through Feature Pyramid Network-inspired fusion and learnable gamma correction. Experiments on the LISU dataset demonstrate superior performance over conventional sequential training approaches, achieving over 30% LPIPS improvement across both AWGN and Rayleigh fading channels, alongside 7% MS-SSIM gains.