Change blindness refers to the inability of our visual system to memorize details in scenery. An example of our brain’s inability becomes obvious when playing the game Spot-the-difference. It’s a game where you have two pictures that look absolutely identical, but in fact small details are different. Often one has to study the pictures for several minutes to discover the well-hidden differences. Steven Le Moan, who until recently worked at NTNU’s Norwegian Color and Visual Computing Laboratory in Gjøvik, wanted to learn more about why humans fail to notice even major changes in images.
Le Moan’s method
According to Le Moan, the human brain can take in the entirety of a picture within microseconds, and leave out the finer details. He claims to imitate the process that happens in our brains when we come across a picture and combine important and unimportant information. So, with the help of his colleagues at the Color and Visual Computing Lab, they developed an image compression method using change blindness. By using an algorithm based on saliency detection and texture synthesis (alternatively referred to as exemplar-based inpainting), he creates a conﬁdence-adjusted saliency map, which he refers to as change blindness map. Pixels with low energy on that map can simply be removed. Therefore large areas of the image can be eventually removed, significantly reducing the file size. It is different from the traditional methods for compressing images like JPEG since it involves understanding the image’s context, not removing random information. This feature makes Le Moan’s method unique because contiguous areas of the image can be removed, without changing what the human eye sees.
The concept these algorithm uses is that the more time someone needs to detect a change or a difference between two images, the less important it is. So it removes parts of the image that it would take a long time to detect since they are the least important. As a result you don’t even notice that they are missing. These pixels can later be recreated from what the surrounding area looks like.
An example of image processing
If someone is given a picture of a cat walking on grass on a sunny day, their first reaction will be to focus on the cat and not on the grass. The computer algorithm can detect which areas of an image are best suited to be removed or altered, which takes advantage of the change blindness phenomenon, without it being noticeable. An unimportant area on the sky can be recreated from how the rest of the remaining sky looks. The grass can be reconstructed to blend in with the rest of the surroundings, despite the differences from the original picture.
What are the limits of image processing?
It is clear that computer algorithms can remove parts of the image they are processing that are considered to have low importance. But how can someone define low importance and what are the limits of the details that can be altered? According to Le Moan, as long as the final picture looks natural there is no problem processing the image. As reported by studies, about 15% of an image can be altered without it looking unnatural and the viewer will not even notice.
Future research in this area includes expanding the field into applications beyond image compression. For example in identification documents to prevent theft or frauds relating to documents, photographs, banknotes or watermarks.
Top image: Reading braille book Stock Image