View All Pages

Class Discussion Summary (Jan 14)

To-Do Date: Jan 18 at 11:59pm

Put the citations to papers discussed here:

Write the Class Discussion Summary Here:

1. Considering that both methods help fix motion blur, is it possible we can combine the two methods to get a more favorable quality picture? How?

Method A: Removing Camera Shake from a Single Photograph - In post processing you predict the motion blur and cancel it out.
Method B: Synthetic Shutter Speed Imaging - Method that takes multiple pictures and takes the average picture.

Yes, they could be combined, since one of them is primarily a hardware approach, and the other is purely software.

However, method A aims to cancel the blur completely, while method B wants to keep the artistic blur. So it would be possible with a slight modification where we take the user input with a subset of the image, i.e. we only apply the method that cancels the blur entirely on the undesired out of focus areas. Thus, the image will not be worse, but there are doubts concerning whether the final image will be improved or not.

Alternatively, we can use the deblurred result from the SSSI approach and replace the sharp image in the blur kernel equation for the first one. This would make the first approach trivial/simpler/less error prone.

2. Which of these methods should be implemented on a mobile phone? Why?

Synthetic Shutter Speed Imaging could be used on modern mobile phones which can take multiple pictures in a buffer; Removing Camera Shake from a Single Photograph is just an algorithm, so it could be used, given the increased processing power of phones.

Since the Removing Camera Shake from a Single Photograph algorithm requires user input, it is probably not appropriate for a mobile phone; the variables are also quite complex to understand and select, hence they are more suited for a dedicated audience and a desktop environment.

The Synthetic Shutter Speed Imaging algorithm can be employed on a mobile phone with additional hardware and no user input, thus it could provide a better quality final result which may please the user.

However, we must consider that mobile phones have cheaper cameras which have the tendency to be oversaturated which might lead to issues with blind deconvolution. Mobile phones also have low computation relatively.

3. In the age of neural networks is this work obsolete? Will NN do better?

Neural networks have come a long way since when the original papers were written; it is possible to use adversarial networks to train a model with original and blurred images to attempt generating the original subject; in order to do so, a huge amount of processing power is needed, therefore the pictures will likely have to be uploaded to a cloud service. Still, it may be doable, and it may provide better results than both proposed solutions.

Typically GANs have been successful in pixel level manipualtion tasks such as deblurring, denoising, image inpainting so it should be successful here.

Also, we think it would be obsolete for the second method but not for the first one because we could still work on blur kernel areas as NN in a learning algorithm. Also, it seems that we have to determine the size of the kernel and categorize the input.

Scribes: Andy Vitek, Alex Salman, Ishaan Paranjape