Metrics for comparing images are essential in the design of image enhancement, compression and modification algorithms. A comparison is needed to give the designer reliable feedback as to the quality of a reconstructed image.
Two of the most popular image quality measures are the root mean square and signal-to-noise ratio. Unfortunately these measures are simple tallies of pixel difference and provide no information about the type of degradation present. Although simple, and consequently provide computational benefits, the RMS and SNR measures cannot meaningfully be applied to images containing text or binary images. Pixel tally measures are also unable to measure perceptual distortion. Take the case of two identical images, one translated one place to the right. These images still appear similar, but an RMS type error will return a large difference.
This Honours project investigated the use of a phase based approach for image comparison. One of the main concerns being that current image comparison measures, such as the RMS, very often do not correlate strongly with human perception as to whether images are similar or different.