Fig. 5: Reconstruction results under combinations of real-world degradations, including mist-induced scattering, platform jitter, signal-dependent noise, and subsampling (sampling ratio: 6.25%).
From: Comprehensive compensation of real-world degradations for robust single-pixel imaging
a, b Are grayscale scenes under varying degradation levels, while c, d show color reconstructions. LPIPS is Learned Perceptual Image Patch Similarity, which evaluates the consistency of high-level semantic features; lower values indicate better semantic similarity between images
