Hybrid Denoising Framework with Deep Learning and Optimization forMixed Noise Removal in Remote Sensing Images

Abstract Remote sensing images (RSIs) are often degraded by mixed noise, particularly salt-and-pepper noise (SPN) and
white Gaussian noise (WGN), which adversely affect visual quality and analytical reliability. Traditional
denoising methods struggle to suppress impulsive artifacts in the presence of mixed noise, especially in RGB
images with varying luminance levels. To address this challenge, a novel hybrid denoising framework is proposed,
integrating spatial filtering with deep learning techniques. The method first employs Contrast Limited Adaptive
Histogram Equalization (CLAHE) to enhances contrast in shadowed and overexposed regions, followed by a
Modified Decision-Based Unsymmetric Trimmed Median Filter (MDBUTMF) to suppress SPN noise while
preserving edge fidelity. Gaussian Curvature Filtering (GCF) then refines geometrically significant regions and a
Denoising Convolutional Neural Network (DnCNN) subsequently recovers fine textures and semantic details lost
during spatial filtering. Additionally, the Adaptive Mayfly Optimization Algorithm (AMOA) is incorporated to
dynamically tunes the filter parameters based on input image luminance variations. Experiments conducted on the
UCM, WHU-RS19 and Landsat datasets demonstrate the superior performance of the proposed hybrid approach,
achieving a peak PSNR of 32.80 dB, an SSIM of 0.888 and a minimum GMSD of 0.0537 under complex mixed
noise conditions. The proposed framework generalizes effectively across diverse image content and noise levels,
making it suitable for a wide range of remote sensing applications.