TITLE:
A Multi-Stage Image Restoration Pipeline through Illumination Enhancement, Environmental Degradation Removal, and Exemplar-Based Inpainting
AUTHORS:
Md. Shamim Imtiaz, Amina Khatun, Abdullah Al Zubaer, Md. Masum Bhuiyan, Sumaita Binte Shorif, Mohammad Shorif Uddin
KEYWORDS:
Contrast Adjustment, Image Restoration, Physics-Based Illumination Modeling, Adaptive Two-Round Search Exemplar Inpainting
JOURNAL NAME:
Journal of Computer and Communications,
Vol.14 No.7,
July
15,
2026
ABSTRACT: Real-world image acquisition under suboptimal environments frequently yields compounding degradations, such as severe under-illumination, sensor noise, atmospheric haze, and dynamic precipitation. These multi-layered artifacts obscure visual features, causing significant drop-offs in downstream computer vision tasks and human interpretation. Traditional image restoration models generally address isolated degradation modes, which often leads to details being washed out, severe noise amplification, or unintended color shifts when exposed to overlapping distortions. To resolve these limitations, this paper introduces a unified, multi-stage restoration architecture that manages illumination corrections, environmental cleanup, and geometric pixel repair within a single execution sequence. The process begins with a lightweight structural detection front-end that evaluates local intensity variance and oriented linear structures to identify the unique degradation factors present in the scene. Based on this diagnosis, the image undergoes systemic illumination correction using a fused pair of complementary gamma functions combined with adaptive histogram equalization to reconcile global and local contrast imbalances. A physics-based absorption light scattering model is subsequently employed to recover fine physical scene radiance from deeply shadowed regions. Following luminance stabilization, deep residual networks conditionally mitigate remaining atmospheric haze, persistent noise patterns, or rain ribbons without stripping structural details or altering clear channels. Finally, an adaptive two-round search exemplar inpainting algorithm bridges gaps from missing data or removed foreground clutter, utilizing boundary color variance thresholds to prevent textural mismatching. Quantitative evaluations on benchmark scenes establish the resilience of the pipeline, yielding an average Peak Signal-to-Noise Ratio (PSNR) of 26.84 dB and a Structural Similarity Index (SSIM) of 0.93.