Closed BenJuan26 closed 6 years ago
BenJuan26
commented
6 years ago Did some profiling on LibRaw::dcraw_process(). Looks like this:
| Function | Duration (ms) | Percentage of total |
|---|---|---|
| raw2image_ex() | 22 | 2.9% |
| scale_colors() | 72 | 9.7% |
| pre_interpolate() | 5 | 0.7% |
| ahd_interpolate() | 498 | 66.8% |
| convert_to_rgb() | 148 | 19.9% |
ahd_interpolate() is only one of many interpolation options, so it might be worth it to check out the other ones to reduce the processing time.
BenJuan26
commented
6 years ago Added this line in getCalibratedImage() to bring the quality down to linear interpolation:
params->user_qual = 0;That reduced the duration of dcraw_process() to an average of 440ms, a reduction of almost exactly 3x from the previous 1326ms. Need to compare how much this affects the final quality of the image. If the quality is satisfactory than the change should be applied. In the future it might be worth proving a user-selectable option for this.
BenJuan26
commented
6 years ago The values for user_qual are documented in LibRaw's API documentation. It could be worth exposing those options in a drop down and setting the variable accordingly, but in my testing linear actually seems to look less artifact-y than the default value. That plus the time savings makes it a no-brainer. Gonna hardcode it to linear for the time being.
A typical call to
processConcurrent()on a single light frame:Reading the image, applying the calibration frames, debayering, white balance, etc. is taking 75% of the processing time. Three times longer than alignment!
Within the call to
getCalibratedImage(), the times look like this:Nearly all of the time is spent in
dcraw_process(), which does the debayering, white balance, gamma, and whatever other processing it needs to do to get the image to its presentable form. So all in all, nearly 64% of the processing time is spent just converting the image from raw to bitmap. That's unacceptable.Since this all resides within an external library, the chances of trying to improve it directly are slim to none. The best thing to do would probably be to find another way to read and process the image. I've made attempts to debayer and white balance with OpenCV (since we're using it already for everything else) but it never seems to turn out quite right.
If another library is used, an essential feature it will need to have is the ability to mess with the Bayer image before processing is done, since that's where we apply the calibration frames. Libraw graciously lets us manipulate the Bayer image directly and then do the processing on the modified version.