16-bit to 32-bit conversion....when?

[MikeWiles]

Just curious, when is the best time to convert images from the 16-bit unsigned integer that comes out of the camera to the 32-bit floating point files that are used when we're processing the image?  I know that in general the answer is going to be "sooner than later", but at what point does it really become advantageous?

When I combine my calibration frames into master frames, should the master frame be saved as a 16-bit or 32-bit?  Since they'll be used to calibrate a light frame that is still a 16-bit image - is it any advantage to have the master frame as a 32-bit floating point?

Similarly, when I calibrate the light frames should they be saved after calibration as a 16-bit or 32-bit?  Because of the division of the flat frame I can see that there might be some loss of precision if it's saved as a 16-bit image.

What about when a stack of images is registered with star alignment?  My sense is that this is the point where it's necessary to save the result as a 32-bit image, but am I correct in that thinking?

Finally, when doing the final integration of the calibrated and registered light frames....that result should be saved as a 32-bit floating point for maximum precision.  But should it be done sooner in the workflow than this?

Thanks,

Mike

[Carlos Milovic]

Any image generated by the integration of frames, should be of 32bits. Intermediate or calibrated image could use just 16bits, but stacked images, always 32.

[georg.viehoever]

Have a look at the lower right of the status bar of the PI application. After integration, it should contain something with "f32", telling you that the integrated image is in float32 format. Of course, you should also save it in this format.
Georg

[MikeWiles]

Any image generated by the integration of frames, should be of 32bits. Intermediate or calibrated image could use just 16bits, but stacked images, always 32.

Thanks Carlos.  My sense was that conversion to 32-bit wasn't truly necessary until you combined...or registered the images.  It's good to get confirmation.  I appreciate it.

Mike

[Andres.Pozo]

IMO the conversion to 32bit should be done in the calibration process. The result of subtracting and dividing by the 32 bit masters can not be stored in 16 bits.

[Carlos Milovic]

Yes, but the noise is larger than the quantization, so I don't see a benefit of using 32bits at that stage.