Many image difference metrics have been developed in the last 4 decades. All of these metrics are constructed to predict perceived image difference, but none have been successful. When we rate image difference we look at different areas in the image, based on the difference in these areas we make a decision of the perceived difference. Information about what draws attention and how we examine images can be used to improve image difference metrics.
This research project investigates the importance of region-of-interest on image d ifferencemetrics. Region-of-interest has been extracted by using an eye tracker, but also by manual marking by the observers. 3 different tasks were performed by the observers while their gaze position was recorded. Further a manual marking of region-of-interest together with a questionnaire to map background knowledge was carried out. The information found on how we perceive and examine images has been applied to different image difference metrics, such as Delta Eab, S-CIELAB, iCAM, SSIM and the hue angle algorithm. The issues regarding how observers look at images given different tasks are also discussed and analyzed.
The results indicate that region-of-interest improves image difference metrics, especially when the metrics already have a low performance in term of linear correlation between perceived and calculated difference. There are no clear evident that one type of region-of-interest outperform other types. The improvement in performance is therefore both scene and metric dependent.
Results also show that observers have different areas of attention according the task given to them, as freeview, rating image difference and marking important regions. The common denominator within every task is faces, and this is clearly important in all tasks for the observers. Within areas of attention will change whether the observer is an expert or non-expert.