Frequency Sub-band Reduction of Spatially Correlated Noise in Images

This study addresses the challenge of mitigating additive spatially correlated noise in images. Our proposed solution involves breaking down the image into frequency sub-bands and individually reducing noise in each segment. By doing so, the noise spectrum within each sub-band becomes more uniform compared to the entire image, enabling the effective use of a neural network originally designed for reducing additive white Gaussian noise. This segmented approach enhances the method's flexibility and adaptability, particularly in handling noise with varying levels of horizontal and vertical correlation. We also investigated the enhancement of noise suppression efficiency through preliminary equalization of noise levels across different frequency sub-bands. Comparative analyses reveal that our method achieves state-of-the-art peak signal-to-noise ratios in processed images, effectively handling both white and spatially correlated Gaussian noise.