09-13-2019, 09:23 AM
There are a few interesting mobile image enhancement techniques which could be very useful if implemented on the PINE64 mobile devices. They could very well improve the decent cameras on the devices to something quite admirable.
The main projects I'd like to see implemented are:
HDRnet - uses downsampling and neural networks to improve image truth, vibrancy, or style. Already has been shown to work on mobile with speed. An open source (Apache 2.0) implementation can be found at https://github.com/mgharbi/hdrnet/ and https://github.com/google/hdrnet/
HDR+ - better automatic RAW development, can align multiple frames and improve color range. I believe it is used on Google phones. An open source (MIT, maybe incomplete) implementation can be found at https://github.com/timothybrooks/hdr-plus
Handheld Multi-frame Super-resolution - merges many images (like those from a live photo) into a higher resolution image. It might be used on Google phones. I don't know of any existing working open source implementation, but this (MIT, very incomplete) may implement it in the future https://github.com/JVision/Handheld-Mult...Resolution
Learning to See in the Dark - intelligent RAW developer which can brighten an image by an extremely large magnitude. I am unsure about the performance, especially on mobile. An open source (MIT) implementation can be found at https://github.com/cchen156/Learning-to-See-in-the-Dark
I mention these projects on the PINE64 forums rather than another forum (like the PostmarketOS or Ubuntu Touch forums) because these projects would likely benefit from firmware enhancements, and splitting the topics across many forums would be a pain to deal with.
I should also mention that I'm in no way a neural network or image processing expert, but I'm sure some other members of the forum are. Maybe this post will catch their interest?
The main projects I'd like to see implemented are:
HDRnet - uses downsampling and neural networks to improve image truth, vibrancy, or style. Already has been shown to work on mobile with speed. An open source (Apache 2.0) implementation can be found at https://github.com/mgharbi/hdrnet/ and https://github.com/google/hdrnet/
HDR+ - better automatic RAW development, can align multiple frames and improve color range. I believe it is used on Google phones. An open source (MIT, maybe incomplete) implementation can be found at https://github.com/timothybrooks/hdr-plus
Handheld Multi-frame Super-resolution - merges many images (like those from a live photo) into a higher resolution image. It might be used on Google phones. I don't know of any existing working open source implementation, but this (MIT, very incomplete) may implement it in the future https://github.com/JVision/Handheld-Mult...Resolution
Learning to See in the Dark - intelligent RAW developer which can brighten an image by an extremely large magnitude. I am unsure about the performance, especially on mobile. An open source (MIT) implementation can be found at https://github.com/cchen156/Learning-to-See-in-the-Dark
I mention these projects on the PINE64 forums rather than another forum (like the PostmarketOS or Ubuntu Touch forums) because these projects would likely benefit from firmware enhancements, and splitting the topics across many forums would be a pain to deal with.
I should also mention that I'm in no way a neural network or image processing expert, but I'm sure some other members of the forum are. Maybe this post will catch their interest?