Compressed domain image matching using symmetric convolution

It is well known that area based image matching based on convolution can be carried out by multiplication in the frequency domain. In compressed JPEG images, we have available the discrete cosine transform coefficients, which are closely related to the frequency domain representation. An operation of symmetric convolution can be defined, which involves extending a sequence by its mirror image before periodically replicating it. This operation can be implemented by multiplication in the DCT domain. We investigate the usefulness of symmetric convolution for image matching where disparities are between zero and one. A simple method is proposed for estimating the location of the maxima of the symmetric convolution, based on the two surrounding points. It was found that the relationship between the estimated disparity and the true disparity was linear, with slope dependent on the individual block. The slope of this relation was found to be uncorrelated with simple properties of the image block. An average slope was computed for five individual images, and also over all images, and used as a correction factor. With this correction factor, accuracies of about 40% RMS error were obtained