Let me assume that the organizations would afford to have high performance super computing to work on high resolution images directly without the need for image down scaling for machine vision applications, what is the compelling need for these super resolution algorithms that down sample and up sample an existing high resolution image?
The need is to use modelling based techniques based on strong mathematical formulations. The second need is to high performance GPU systems, accessible in every research center and even good prices for the access at homes as well.
Dear Sivaramakrishnan Rajaraman,
Look the link, maybe useful.
https://ieeexplore.ieee.org/abstract/document/8451496
https://books.google.az/books?id=upooPUQLIA0C&pg=PT309&lpg=PT309&dq=What+is+the+compelling+need+for+image+super+resolution?&source=bl&ots=3BEUf0NWdS&sig=ACfU3U0DzcgEzFWEL1n_FYI5jyKKdkznwA&hl=ru&sa=X&ved=2ahUKEwjG4LHD0r7mAhVhyoUKHflvDEwQ6AEwC3oECAoQAQ#v=onepage&q=What%20is%20the%20compelling%20need%20for%20image%20super%20resolution%3F&f=false
Regards,
Shafagat
To improve the speed of other underlying algorithms and/or make algorithms scale well to high-dimensional data / big data.
I have mentioned that i have enough computing power to work with high dimensional data straight away. Then what would be the real need to use these algorithms?
Extracting extra features at different scale levels (building a feature pyramid), for example it could help in generalization of object detection in the wild.
Compression. A lot of video compression algorithms use YUV4:2:0 colour space. YUV4:2:0 means that the black and white channel (Y) has four times the number of pixels as the colour channels (U and V). You save a lot of bandwidth by losing some colour that humans can't really perceive.
Having an effective and fast super resolution algorithm (which is accurate enough to fool humans) would let images (maybe video) be compressed very small for transport and then enlarged accurately at the other end.
Pamela Johnston
He assumes that HPC is attainable, thus having “fast processing” algorithms is not their objective, see my answer above.
Abbas Cheddad
Yes, but datasets (or, rather, images for any purpose) still need to be stored and transported over a network (and provided in an accessible way if they're going to be shared). Super resolution is an end to itself in the field of compression, and it may well outperform the traditional frequency/block-based approaches (e.g. JPEG or H.264/AVC, HEVC). If that's the case, the internet might become filled with low resolution "thumbnail" images that are simply enhanced post-download by a browser (saving on bandwidth/download times/storage requirements for Instagram etc). That's a (albeit generic!) compelling need for super-resolution (i.e. downscaling and then upscaling), and some of the super resolution papers cite that as their motivation.
Learning features from super resolution images, as you suggest, may well be useful for future machine vision algorithms. It will certainly be useful if super resolution becomes adopted as a common means of compression!
Pamela Johnston
I agree, and I did mention in my first post the "speed of processing" of algorithms but that does not seem to be of interest to the creator of this post based on his reply (maybe my answer was understood as merely for internal processing). When working with CODEC and network communication, speed of processing (i.e., down sampling / super resolution) becomes apparent for multimedia communication because of the bandwidth (as you mentioned) and as you cannot guarantee the client is as lucky as you are in having access to HPC.
Thanks for the awesome explanation Pamela Johnston and Abbas Cheddad
Sivaramakrishnan Rajaraman
You are welcome! Wishing you all the best in your research.
Happy new year everyone!
Abbas Cheddad and Pamela Johnston Many thanks once again. Apart from the compression factor which is a typical computer vision requirement , what do you think is the value of super-resolution Deep Learning architectures for medical image analysis?
In general, I would be extra-careful when dealing with medical image analysis for the sensitivity of the domain. However, to answer your question, I would think of the removal of artifacts generated through up-sampling, increasing the resolution of projections acquired using devices with narrow field of view (FOV), or multi-modal image registration/matching etc.
I think this situation is useful in some special applications only such as in medical because the uncompressed high-resolution images need huge storage space, supercomputers and very expensive
- Badges
- Science topic
- Similar topics
- Computer Science
- Computing