Evaluating the Performance of Image Super-Resolution Baselines in the Presence of Spatially Correlated Pink Noise

Single-image super-resolution (SISR) frameworks assume that image degradation is caused by independent Additive White Gaussian Noise (AWGN). However, physical camera sensors, atmospheric scattering, and imaging hardware often generate spatially correlated, low-frequency colored noise patterns. This study sought to answer how classical digital upscaling and linear filtering baselines perform when restoring images obscured by non-white noise. We hypothesized that standard bicubic upscaling would fail to accurately reconstruct high-resolution targets because it cannot mathematically separate cloudy noise clusters from genuine edge data. To test this, high-resolution benchmark targets were downsampled by a scale factor of 2, corrupted with simulated 1/𝑓 pink noise at three distinct intensities (σ= 0.10, 0.20, 0.30), and reconstructed using standard bicubic upscaling and spatial Gaussian low-pass filtering. The quantitative results, measured via Mean Squared Error (MSE), demonstrated that our initial hypothesis was rejected; standard bicubic upscaling consistently maintained a lower mathematical error index than the filtering method. While the spatial Gaussian filter changed the visual output by removing the cloudy noise, it showed edge blurring of sharp edges. This uniform smoothing across the entire pixel array created a mathematical error penalty that exceeded the benefits of background noise reduction, revealing a compromise in linear image restoration.