Subjective evaluations of example-based, total variation, and joint regularization for image processing

We present subjective evaluations of example-based regualrization, total variation regularization, and a proposed joint example-based and total variation regularization for image estimation problems. We focus on the noisy deblurring problem, which generalizes image superresolution and denoising. Controlled subjective experiments show that the proposed joint regularization can yield significant improvement over only using total variation or example-based regularization, particularly when the example images contain similar structural elements as the test image. We also investigate whether the regularization parameters can be trained by cross-validation, and the difference in cross-validation judgments made by humans or by fully automatic image quality metrics. Experiments show that of five image quality metrics tested, the structural similarity index (SSIM) correlates best with human judgement of image quality, and can be profitably used to cross-validate regularization parameters. However, there is a significant quality gap depending on whether the parameters are cross-validated by humans or with the best image quality metric.