Class-adapted blind image deblurring (by Marina)

In the previous post, I introduced the concept of blind image deblurring: the problem when we have an observed blurred image and we want to find an underlying sharp image and a blurring filter.  Also, I mentioned that there are different types of blurring filters and they have different characteristics. What I did not mention is that there are also different types of sharp images (figure 1).

Figure 1: Different types of images.

Different types of sharp images can have completely different structure. For example, look at a (document like) text image which has sharp transitions and usually comes in two tones, black and white. On the other side, a natural image has smooth regions (like sky or lake), texture (bridge, trees) and some sharp edges (boat). An image of a face can have small texture (like area below the eyes), that is very difficult to estimate ones when is blurred.

As we have a problem where we need to estimate two unknowns (an underlying sharp image and a blurring filter) from only one observation (a blurred image), we usually introduce some kind of prior knowledge on both the sharp image and the blurring filter. A prior knowledge of image can represent a characteristic or a set of characteristics about sharp images itself. The similar is for blurring filters. But, as you know so far, there are different types of images and different types of blurs, so how to find a unique prior knowledge (one for an image and one for a blur) that can cover these differences?

We proposed one solution in the paper presented at the International Conference on Image Processing (ICIP) 2017. The main idea behind the proposal is that we can learn an image prior for different types of images. For example, if we know that we have a blurred image of text, we can use sharp images that contain text (other images) to learn some underlying characteristics of text images.

Figure 2: Clean image of text, blurred image of text and our estimated image.

More information can be found here: https://arxiv.org/abs/1709.01710