Ternary-Decoupled Latent Diffusion for Low-Light Enhancement of Hydropower Inspection Images

Abstract: To address the loss of key textures such as micro-cracks and corrosion when enhancing inspection images captured in the underground powerhouses of hydropower stations under weak, uneven illumination and strong noise, a texture-enhanced latent diffusion network (Tex-LDM) is proposed. The low-light image is decoupled into illumination, base-color and texture priors. A ternary prior extraction network learns the supervised priors, three conditional latent diffusion branches with independent parameters estimate the distributions of the high-quality priors, and a texture-injection Transformer fuses the texture prior into brightness recovery through cross-attention, so that noise suppression and texture reconstruction are accomplished jointly. On the self-built hydropower inspection dataset Hydro-Dark, the peak signal-to-noise ratio (PSNR) is 2.30 dB higher than that of Reti-Diff. On the public LOL-v1, LOL-v2 and SID datasets, the Fréchet inception distance (FID) and the blind image quality score are the lowest among the compared methods, and the FID on LOL-v2-synthetic is 42.45% lower than that of Retinexformer. The results show that ternary decoupling and texture diffusion priors alleviate texture-noise confusion under extremely low signal-to-noise ratios, suppressing dark-region noise while preserving key defect textures.