Blind Deconvolution:

Image Restoration vs. Image Enhancement

 

Image Processing can be applied for one of two basic reasons, image restoration or image enhancement. The goal of restoration is to achieve an accurate depiction of the scene being imaged. In contrast, enhancement strives to create the most visually appealing image. An example of image enhancement is to slightly blur an image to reduce the amount of visible noise. There is a gray area in this definition. An analyst might apply an edge detector to make it easier to count interference fringes even though the processed image is less accurate than the original. The image itself might be less accurate, however the information the analyst gathers from the image is more accurate.

 

Deconvolution is used to achieve either or both of these objectives. Blur removal typically makes the image more visually pleasing while creating a replication that is more accurate. Blind deconvolution is also a very useful image enhancement tool; its use in restoration comes with a caveat. How can you verify the accuracy of the method if some portion of the input is based on a guess? Aside from special situations, generally the answer is that you cannot directly determine the accuracy of your result. So does this uncertainty preclude the use of blind deconvolution for image restoration?  Based on the increasing amount of scientific applications that use blind deconvolutions, this answer is no.

 

The reasons are: In order to produce a positive result, a very good guess is required. For iterative techniques, poor input produces non-converging results, i.e. the algorithm does not realize it is straying further and further away from the right answer. Poor input into non-iterative methods produces results that are dominated by image artifacts. Blind deconvolutions are mostly applied to images where the blur function is space-invariant. This means that we expect to see the same blur function occur at any point in the image. If artifacts do occur, they are not isolated events within the image. The analyst will quickly identify non-real objects in most images. Edge ringing is the most common artifact that occurs, and this can found using either qualitative or quantitative methods. Suspect features in a processed image can also be compared to the original image. Although blurred, these features, if real, must be in the original image. Except in cases of extreme blur, the analyst will easily locate the feature in the original image.

 

The image analyst has a good understanding of the imaging system and the object that the camera is imaging. For example, planetary probes, such as NASA's Cassini, send back images of places we have never seen before. Even though he/she possesses little knowledge of the planet or moon, the analyst expects to see a landscape that has at least similar characteristics to those that he has seen before. The analyst then draws on his knowledge of previous landscapes to process the new images. The human mind sees very accurate depictions of scenes all the time, and is well practiced by nature in the art of determining image quality. It is this knowledge that provides an additional quantitative improvement to image restoration.

 

The utility of SeDDaRA for image restoration depends on these qualities.  The analyst is asked to choose a reference image that resembles the spatial frequency content of the original, process the image with a few chosen parameters, and then evaluate the result.  In absence of standard image metrics, the analyst determines whether the result is adequate or needs further tuning.  However, the speed of SeDDaRA enables the analyst to perform this task in a few minutes as opposed to hours. On this page, we have expressed our opinion on the use of blind deconvolution for image restoration.  We are obviously in favor of it.  However, others may not be.  If you have a concern about how this argument was presented or about the content, please let us know.