Bi-linear Image Resample

[EricLarsen]

I've got an application that is processing some fairly large images. They are typically 4800x1280 pixels, 32 bit images. As part of the processing, I resample the image to adjust the aspect ratio. The IMAQ Resample (bi-linear option) function is fast at this, typically around 50-60 ms on my development computer.

 

Unfortunately, this application is going to be distributed to multiple customer sites, and the IMAQ Resample is part of the Vision Development Module, which requires a run-time license for each client computer. That's not acceptable to our customer, especially since resampling is the only Vision Development function I use, the rest are custom fuctions.

 

To get around this I wrote my own solution in G. It uses the Resample Waveform.vi function to do a row-by-row, then colum-by-column resample. It works, but it's fairly slow, around 3-4 seconds (sometimes up to 10s, depending on what mood my computer is in). Not bad, but I'd like to do better. Admittedly I haven't tried to optimize it yet, it's not a high priority.

 

So before I dive into this, has anybody got some code or link to an efficent algorithm they would be willing to share that can do this?

 

[GregSands]

Have you looked at using the built-in Interpolate 2D routine from the Mathematics/Interpolation palette? Dive into that and rework the core to hard-code as much as possible (if your parameters are all constant).  What is the resulting image size?

[EricLarsen]

I busted out my 30 year old copy of Numerical Recipes and programmed up a quick 2D bilinear routine that works pretty well. I got the resample time down to ~300 ms. Still not as good as the IMAQ function, but good enough.

 

I had looked at the built in 2D interpolate routine in the past, and for reasons I can't remember I'd rejected it. I took another look at it today and didn't get very good results (~15s). It does some data tweaking and massaging that slows it down. But the core interpolation code looks exactly like the Numerical Recipes code, so I doubt it can do much better than mine.