Given enough time, a digital camera could take a dozen well-exposed photos, and software could stitch them into a perfectly focused composite. But if the scene is changing, or if the photographer is trying to hold the camera steady by hand, there may not be time for a dozen photos. When time is short, says postdoc Sam Hasinoff, lead author on the paper, "there's a trade-off between blur, on the one hand — not having an image which is in focus — and noise, on the other. If you take an image really fast, it's really dark; it's not going to be of high quality."
Hasinoff, MIT professors Fredo Durand and William Freeman, and Kiriakos Kutulakos of the University of Toronto devised a mathematical model that determines how many exposures will yield the sharpest image given a time limit, a focal distance, and a light-meter reading. Hasinoff says that experiments in the lab, where the number and duration of digital-camera exposures were controlled by laptop, bore out the model's predictions.
A digital camera could easily store a table that specifies the ideal number of exposures for any set of circumstances, Hasinoff says, and the camera could have a distinct operational setting that invokes the table. The multiple-exposure approach, he says, offers particular advantages in low light or when the scene covers a large range of distances.
For the time being, however, the technique is limited by the speed of camera sensors. Today's fastest consumer cameras can capture about 60 images in a second, Hasinoff says. If the MIT researchers' model determined that, under certain conditions, the ideal number of exposures in a tenth of a second would be eight, the fastest cameras could manage only six. "But there's still a big gain to be had," Hasinoff says.
The Graphics Group's work on multiple-exposure composites uses an analytical approach first presented at this summer's Siggraph — the major conference in the field of computer graphics. There, Anat Levin, who was a postdoc at the time, Durand, Freeman, and colleagues described their "lattice-focal lens," an ordinary lens filter with what look like 12 tiny boxes of different heights clustered at its center. Each box is in fact a lens with a different focal length, which projects an image onto a different part of the camera's sensor. The raw image would look like gobbledygook, but the same type of algorithm that can combine multiple exposures into a coherent composite can also recover a regular photo from the raw image.
"Only time will tell whether that new, proposed piece of hardware will be better than the others, but I think their way of analyzing the whole thing is brilliant," says Marc Levoy, a professor of computer science and electrical engineering at Stanford University. "There's been a lot of work on different ways of extending the depth of field, and what this paper did was, it tried to analyze all of them together. And I actually think that it's a seminal paper. I think it's a landmark paper."