A New Preprocessing Method for Measuring Image Visual Quality Robust to Rotation and Spatial Shifts.

Measuring the visual quality of an image is an extremely important task in computer vision. In this paper, we perform 2D fast Fourier transform (FFT) to both test and reference images and take the logarithm of their spectra. We convert both log spectra images to polar coordinate system from cartesian coordinate system and use FFT to extract features that are invariant to translation and rotation. We apply the existing structural similarity (SSIM) index to the two invariant feature images, where no extra inverse transform is needed. Experimental results show that our proposed preprocessing method, when combined with the mean SSIM (MSSIM), performs better than the standard MSSIM significantly in terms of visual quality scores even when no distortions are introduced to the images in the LIVE Image Quality Assessment Database Release 2. In addition, when images are distorted by small spatial shifts and rotations, our new preprocessing step combined with MSSIM still performs better than the standard MSSIM.