Realtime image registration/stack (17 frames/sec)
Posted 26 July 2014 - 05:55 PM
Posted 31 July 2014 - 12:34 AM
Their GPU has three times the GPU cores I'm using too! The quadro has 1536 cores vs my AMD's 490 and it also has full PCI-E bandwidth, the apple GPU integration only has a couple of PCI-E express channels!
I've been doing some work on the Lanczos - specifically around upscaling into images as little as 60x60 pixels in association drizzling. I'll need to re-optimise after but it's still going fast enough
Posted 10 August 2014 - 12:26 PM
I've been dogfooding today - plenty of thoughts and ideas, however the testing of the focusing seems to be doing a good job. The green dots are the frame quality, the band of green marks the band between highest and lowest grades. Bottom is 0%, top being 100% fine detail - so in theory it will never get there
This was shot through the conservatory roof - the weather outside was very bad this morning and is still exceptionally windy (UK had national weather warning for heavy rain and winds!):
Screen_Shot_2014_08_10_at_14_40_58_Preview01.png 246.93KB 1 downloads
I have a couple more ideas on this - such as exposure adjustment to maximise detail - also I like the HDR idea however for now I need to get some heavy testing done as it works but not completely to my satisfaction (I had difficult conditions with clouds!).
Cropped screenshot - showing higher quality grading during a downpour (this was through the rain soaked conservatory window) - the titan is only 659x495 pixels:
Screen_Shot_2014_08_10_at_10_59_08_Preview011.png 245.43KB 2 downloads
Not bad.. so I may add gradient histograph too - but for now.. that works. Time to look at the app and get things tested to destruction - just waiting for the sun set - it's meant to be clear tonight
Posted 23 January 2015 - 07:09 AM
I thought I'd give an update on this.
I have continued to work on this in the background, including porting the pipeline to OpenCL rather than Apple's own form of it.
So far the realtime pipeline supports:
* Rotation & Translational alignment using FFTs - this works for DSO, planets, lunar and solar.
* Drizzle stacking with upsizing and optional regional selection
* Focusing using image sharpness rather than FHWM (as above)
* Darks/Flat bias subtraction
* Lucy Richardson deconvolution using 2D PSF by FFT
* Lucy Richardson deconvolution using symmetrical 2D PSF by IIR - the system curve fits the impulse required. It's stupidly quick.
* Histogram stretches
* OpenGL rendering
I started LR non-symetrical 2D IIR research, however I think the complex maths will result in a non-realtime. At this point my MBP's GPU failed (it's one of the 2011 15" MBPs that is known to have a problem.. yet another Apple GPU-gate) and leaving me with no OpenCL 1.1+ capable machine... as I was then made redundant once I get a new job.. I'll get a new machine with a GPU that will allow me to continue..
Here's testing of the LR IIR deconvolution using a generated 2D PSF. Deconvolution/Convolution with unoptimised code in GPU for the lenna test picture below takes 47 microseconds. (i.e. 47x10E-6 = 0.00047 seconds). In the picture - left is deconvoluted, middle original, and right convoluted (i.e. using the same process but inverse of sharpening).
Edited by NickK-UK, 23 January 2015 - 07:10 AM.
- Relativist likes this