This week we were able to present our proposal slides to our peers, Prof. Sullivan, and our TA. It allowed us to share our ideas, plan of action, and next steps as a team. During this presentation, we encountered some questions concerning our CNN approach. We think this might be the biggest risk because we want to make sure out photo processing algorithms are effectively able to deblur. We plan to use a encoder-decoder style CNN, which will encode the input image into a low level embedding which will be led to a decoder to get our desired deblurred output. In order to mitigate this risk we hope to do more research on this CNN approach. Since this is the core of our project we need to make sure the approach we take is the most efficient. We also intend to read into lectures and paper about deblurring techniques to give us a clearer understanding of what we should expect. In terms of our contingency plan we intend to research more about GAN networks.
Aside from the presentation, we also met with fellow signals and systems professor Aswin Sankaranarayanan to discuss some of the project requirements and CNN approach. Aswin stressed that we continue to refine our exact problem statement, he suggested that we consider one kind of motion blur because it may be too difficult to create and train a network that can identify and deblur multiple types of blur. So we did some discussion and we are going to primarily focus on the blur that comes from camera shake (spatially-invariant motion blur). Only focusing on this kind of motion significantly reduces the size of the problem space which will narrow our scope and give us a more achievable goal when it comes to training our CNN. Another risk that Aswin mentioned was that determining what frame is considered “blurry” is a very difficult problem because there can be blur from many different sources. So for example if we want to classify a frame as blurry, for our system what we consider blurry is blur from camera shake, whereas there may be many other sources of blur such as focal length, the scene and exposure time. This makes our classification process more complicated because we only want to detect one particular kind of motion blur. Good thing we have not started implementing anything and are still in the design phase of this process.
In terms of software packages we have started looking into how we can utilize OpenCV to handle images. This includes how we want to load them in with our Python scripts and access multiple images. While further distribution of these images is still not known (for example seeing if they’re blurry and handing them to the CNN) initial progress can begin for basic scripting to handle how we imagine this handoff will look in the short term. While this is very rudimentary as of right now, progress is being made on figuring out how to use it/background research is being conducted but more progress is expected to be made next week as much of our time was spent working on our presentation and making sure we were prepared for that.
0 Comments