6/17/2015 Daniel Dexter, ECE ILLINOIS
Written by Daniel Dexter, ECE ILLINOIS
Graduate students Jia-Bin Huang and Abhishek Singh took a new approach for solving a 2-D image super-resolution problem by looking at it from a 3-D perspective.
Their discovery presents an elegant solution to this inverse problem, which has plagued researchers in the imaging field for years.
“If you take an image that is too far away, you don’t get enough resolution, or if your hand shakes as you take a picture, it could be blurred,” Huang said. “The inverse problem is when you want to estimate what the high-quality image would have been from what you got. This work is about super resolution, in which you have a low-resolution image, but you want to uncover the invisible details inside it.”
Their work on the subject has earned an opportunity to present their findings at the prestigious Computer Vision and Pattern Recognition (CVPR) conference in June. The paper, titled “Single Image Super-Resolution from Transformed Self-Exemplars,” was selected for an oral presentation, an honor that around five percent of papers receive.
Prior to their discovery, researchers often would use examples from external data to super-resolve pictures. This involved collecting large-scale datasets of natural scenes in order to map out and replace the low-resolution aspects of the image. This method is limited because of the difficulty in finding relevant images in an external dataset.
Through their algorithm, the examples are taken from within the image and reoriented to align with the 3-D scene. This creates a true perspective at which the image is being viewed, and doesn’t require collecting external data. This algorithm, as their research shows, not only produces crisper images, it is also much simpler to apply than the competing one.
“The transformation allows the ability to find hidden, fine details of the image that may otherwise be obscured by the changes that accompany viewing from different directions,” Huang’s and Singh’s advisor Professor Narendra Ahuja said. “For example, if you look at a wall straight on, all the bricks are visible to the same degree, but if you look at the same wall from near one of its ends, the far away bricks appear to be increasingly distorted and compressed. The transformation neutralizes this effect -- a simple idea that makes a big difference in the quality of the super-resolved images.”
Huang sees this method having a variety of applications including in surveillance functions, like revealing license plate numbers that were difficult to see, for example. He said Ahuja encourages his students to aim at fundamental research that also helps solve important problems, instead of limiting attention to incremental improvements in already existing work.
Huang has taken that message to heart in his research as he takes on his next challenge of finding a way to effectively de-blur an image. However, until then, he is looking forward to presenting the group’s findings at CVPR and introducing the field to a new way of thinking about super-resolution.
“I think that our approach is unique in the sense that it treats images not just as 2-D signals,” Huang said. “Instead, we treat them as 2-D projections of a 3-D scene and use that information to help solve inverse problems.”