Hybrid Learning-Based Multi-Domain Enhancement Framework for Heterogeneous Image Modalities

Abstract

Medical and environmental images such as MRI, CT, radiographs, retinal fundus photographs, satellite, aerial and hyperspectral scenes are routinely degraded by noise, compressed contrast, haze and uneven illumination. Classical enhancement operators are fast and interpretable but need image-specific parameters, while deep networks are data-hungry, opaque and brittle across modalities. This paper proposes a Hybrid Learning-based Multi-Domain Enhancement (HLMDE) framework that keeps a transparent chain of transform-domain (BM3D) denoising, log-Fourier homomorphic correction, DFT high-frequency emphasis, CLAHE blending and gamma mapping, and learns per image the six parameters that drive it. A 24-dimensional modality-aware descriptor is projected by PCA and clustered by K-means into unsupervised appearance domains; a random forest predicts the relative quality of 82 candidate parameter vectors and selects the best; cluster-balanced augmentation and fitted-Q-iteration reinforcement learning are evaluated as further refinements. On a benchmark of 694 real images from seven modalities and six public sources, with physically motivated degradations, HLMDE raised mean fidelity from 15.22 dB / 0.529 SSIM to 26.44 dB / 0.813 SSIM. It outperformed CLAHE, histogram equalisation, wavelet-CLAHE, fixed-parameter hybrids and an end-to-end patch MLP (+2.80 dB, p < 10?²¹), and matched a strong BM3D baseline overall while exceeding it on Landsat (+1.51 dB), CT and hyperspectral data. Predicting candidate quality instead of regressing parameters contributed +1.05 dB (p < 10??), whereas augmentation and RL refinement gave no significant gain. In leave-one-modality-out tests the learned controller lost 0.75 dB on unseen modalities against 2.14 dB for the MLP. The framework runs on a CPU and returns an auditable parameter log for every image.