The conventional application of Delta E to images is indeed "a start" but it doesn't account for near neighbor relationships. The visual impact on the image can be dramatically different with the same delta E outcome. That's why I also use the I* metric which tracks changes in both image lightness and image contrast.
Indeed, but at least there's some validity to higher deltaE predictions and results versus lower ones by such analysis. You get an average deltaE over 1 and change when comparing the target patches used to build the profile for this analysis, something isn't kosher.