The question I have is related to image quality evaluation. Suppose you are supposed to perform OCR (Optical Character Recognition) on a set of document images. Some images, however, are very blurring, and for these images blind deconvolution is needed. However, there are good images that do not need deconvolution. Then my question is: could I find a criterion to automatically separate these images into two groups: blurred image and clear images and then treat them differently when performing OCR?
Yes, but doing so is difficult. You are basically assuming that when a certain feature is sufficiently present in an image, you can apply a filter and realize an improvement in OCR accuracy. This implies that you can find the right feature, apply the right filter, and evaluate OCR accuracy correctly. Since there are lots of features that might be relevant (edge prominence, high-frequency energy, PSF estimation, unfiltered image OCR accuracy metrics (as @Matt mentions), etc.), lots of filters available (blind deconvolution, among others), and you'll need lots of ground truth images to prove system effectiveness, you are looking at a lot of work and experimentation to create a system that works well.
You also need to make sure you are not working against the OCR software itself, which may do its own image processing to improve its accuracy. In other words, make sure that artifacts introduced by your deblurring do not make OCR accuracy worse, even if the deblurring was done "well" from a deblurring perspective.
All that being said, it should be possible to do this. You might even be able to do this adaptively, by automatically and iteratively adjusting parameters and approaches to optimize the result.