Super-resolution Image Reconstruction Sina Jahanbin Richard Naething EE381K-14 May 3, 2005
Summary of Super-resolution Results in Literature Subjective results most prevalent reporting method Many papers lack implementation complexity information
Recursive Least Square SR Method [Kim et al., 1990] SR Image LR Image PSNR (dB) 15.5360 16.8558 MSE 0.0239 0.0206 SSIM 0.6134 0.4630 Original Image Under-sampled Noisy Image … First Iteration Second Iteration Final SR Image
Wavelet Based Super-resolution [Bose et al., 2004] SR Image LR Image PSNR (dB) 30.8801 15.7670 MSE 8.1272e-004 0.0107 SSIM 0.8709 0.4391 Original LR Noisy SR Image
“Structural SIMilarity (SSIM)” [Wang et al., 2004] SSIM is an improved version of the Universal Quality Index mention in class Other perceptual models have been based on MSE, but with error weighted based on visibility Error visibility versus loss of quality? Problems with quantifying loss of quality Multiplicative noise Source: Image Quality Assessment: From Error Visibility to Structural Similarity [Wang et al., 2004]