Anisotropic Total Generalized Variation Enhanced Deep Image Prior for Image Denoising
Jue Wang, Jianlou Xu, Yan Hao, Limei Huo, Zengbo Wang, Bohan Li · Symmetry · 2026
To enhance the performance of deep image prior, we propose a novel image denoising model that embeds an anisotropic diffusion tensor into the total generalized variation model and combines it with the deep image prior. The proposed tensor weights deep gradients and guides gradient orientation, which effectively preserves sharp edges. We solve the corresponding minimization problem using the augmented Lagrangian method and the alternating direction method of multipliers. Experimental results show that the proposed method can remove noise while suppressing staircase artifacts and enhancing edge structures, yielding restored images with clearer edge details. Both quantitative metrics and visual comparisons show consistent improvements over competing methods across multiple noise levels, with more pronounced advantages in edge preservation.