Journal of Pattern Recognition Research

There are a number of distortions in image acquisition, processing, compression, storage, transmission, and reproduction. Existing image metrics provide a good judgement on these image distortions. However, they may fail to measure the distortions accurately due to the misalignment between the test image and the reference image. In this paper, we proposed the use of a pre-processing method so that the processed image is robust to the affine transform, which can be treated as a combination of different kinds of deformations, e.g., translation, rotation, scaling, and skewing. We start by extracting the scale invariant feature transform (SIFT) features from the reference image x and the distorted image y, and then find the matching key points between the two images. Next, we solve for the parameters of the global affine transform, and finally, we compensate the image y by an inverse affine transform to generate an image y * . We can then use standard metrics on (x,y * ) to measure the image visual quality. The metrics we study in this paper include PSNR, MSSIM, and VIF. Experimental results show that this pre-processing method improves the metric scores considerably.