ABSTRACT The increasing demand of imaging systems in urban surveillance and medical diagnostic systems is the ability to adapt to non‐homogeneous acquisition conditions without manually adjusting the system parameters. Current methods are generally tailored to specific domains, struggle to generalise to unseen noise patterns, and often depend on retraining computationally intensive deep networks that require large amounts of annotated data. We seal this divide with MDAF++, a single filtering framework that automatically classifies the input domain, and dynamically repurposes its processing strategy at inference time. Instead of parameterized models, or trained weights, MDAF++ interconnects wavelet‐based noise estimation, retinex‐based illumination normalization, multi‐scale frequency fusion, and a dedicated vessel‐preservation branch of retinal data. We test the framework on four standard benchmarks, ExDark and LOL‐v2 to test urban low‐light imagery, DRIVE and STARE to test retinal fundus imaging, and find it is consistent in preserving diagnostically and operationally critical structures. On DRIVE, MDAF++ can achieve a vessel sensitivity of 0.71, which is significantly better than the best classical baseline (0.49), and provides perceptually superior reconstructions in low‐light scenes. These findings indicate a conscious design trade‐off: structural and perceptual utility are put at the forefront over pixel‐wise fidelity, and all the differences reported are statistically significant ( p < 0.001). MDAF++ is currently capable of running on edge platforms where it is infeasible to execute the inference using a GPU. The framework thus provides a practical, cross‐domain denoising pipeline which emphasizes anatomical integrity and scene clarity over traditional fidelity measures.
Haque et al. (Mon,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: