<p>A major challenge in contemporary imaging, especially in low-light microscopy, is the removal of mixed Poisson and Gaussian noise. Interdependent noise components complicate the separation process and demand precise methodologies. Existing deep learning and model-based approaches frequently fail to capture complex dependencies or require extensive labeled datasets, limiting their applicability in practical scenarios. Additionally, many methods address only a single noise type or rely on oversimplified assumptions, which diminishes their effectiveness. As a result, previous approaches are unable to address mixed noise types and intricate dependency structures without large labeled datasets. To address these challenges, this work introduces a two-stage methodology. The first stage employs directional total variation (DTV) to suppress Gaussian noise. In the second stage, Poisson noise and image details are separated using copula density functions. The approach leverages semiparametric copula models, which combine a parametric copula with nonparametric marginals, enabling flexible and precise noise separation. Quantitative evaluations on synthetic and real images demonstrate that the proposed approach improves the structural similarity (SSIM) by up to 4–5% and the peak signal-to-noise ratio (PSNR) by about 0.8–1.9&#xa0;dB compared with the Bilateral Poisson-Gaussian Separation (BPGS) and related variational baselines. The method also preserves edge sharpness and fine texture more effectively than conventional total variation or plug-and-play deep priors. Beyond imaging, the dependency-aware formulation of the proposed framework establishes connections with blind source separation (BSS) problems in other domains, such as biomedical signal analysis and multimodal fusion, highlighting its cross-domain potential. Overall, the findings indicate that the copula-based method is robust in addressing complex mixed-noise challenges encountered in practical image restoration.</p>