December 21, 2010

“Zoom in. Now… enhance.” – results

by Andrey Filippov

UPDATE: The latest version of the page for comparing the results.

This is a quick update to the Zoom in. Now… enhance. – a practical implementation of the aberration measurement and correction in a digital camera post published last month. It had many illustrations of the image post-processing steps, but lacked the most important the real-life examples of the processed images. At that time we just did not have such images, we also had to find out a way to acquire calibration images at the distance that can be considered “infinity” for the lenses – the first images used a shorter distance of just 2.25m between the camera and the target, the target size was limited by the size of our office wall. Since that we improved software combining of the partial calibration images, software was converted to multi-threaded to increase performance (using all the 8 threads in the 4-core Intel i7 CPU resulted in approximately 5.5 times faster processing) and we were able to calibrate the two actual Elphel Eyesis cameras (only 8 lenses around, top fisheye is not done yet).  It was possible to apply recent calibration data (here is a set of calibration files for one of the 8 channels) to the images we acquired before the software was finished.

So here they are – the images processed with ImageJ-Elphel plugin to the ImageJ program as described in the previous blog post: There are multiple parameters that influence the correction results, and generally as the aberration compensation  increases, images get “sharper”, but more of the processing (and initial compression) artifacts become more visible, so the table below includes image sets with different correction amounts. Links in the “Correction” columns (“medium”, “low” and “high”) lead to the comparison pages, where you can switch between “raw” and “enhanced” images of the 8-image sets. Left side of the page shows the image area in “actual pixels” mode (zoomed 100%), on the right side there is a navigation thumbnail followed by 8 of the camera channels selection thumbnails. It is possible to move around the main image by dragging either it or the rectangle on the navigation thumbnail. Double-click on the main image  switches between raw and processed images in a pair.

The originals were acquired in JP4 format, combined by 3, as Elphel Eyesis camera multiplexes images from each 3 sensors to the same processor board. The “Raw JP4/info” column includes links to these 3 original images through the online Jeffrey’s Exif Viewer – it show image metadata, and for the Eyesis channel 1 (it has the GPS attached) it includes the the map showing the location where the image was acquired (channels 2 and 3 do not have geo data), it also has a link to the same location on on Google Maps. Calibration image compression quality (image used to calculate an array of the  optical PSF ) is even more important than the compression quality of the normal images, and with the camera #2 we unfortunately realized that our regular JPEG quality of 95% is not enough, only after that camera was already shipped, so the correction results are limited by the quality of the PSF measurements we’ve made. Eyesis camera number 3 is still in our possession so we re-measured the calibration data in virtually lossless mode (no quantization) with JPEG quality of 100%.

When comparing image pairs you can see that there is more difference in the details sharpness in the off-center areas, consistent with the original goals of this work. In the center areas lenses have sufficiently high resolution without additional processing.

Aberration correction results table

Camera number Correction amount Raw JP4/info Calibration
image
quality
Comments
2 medium 1+GPS 2 3 95% Salt Lake City, Main St., Gallivan Plaza
2 medium 1+GPS 2 3 95% Inside hardware store
2 medium 1+GPS 2 3 95% Inside hardware store
3 medium 1+GPS 2 3 95% Salt Lake City, Main St., Gallivan Plaza. One of the camera channels (number 1/3) was severely out of focus during both acquisition and calibration (problem fixed for other images)
3 low high 1+GPS 2 3 100% Salt Lake City, Main St./NorthTemple St.
3 low 1+GPS 2 3 100% Salt Lake City, Main St., Gallivan Plaza.
3 low high 1+GPS 2 3 100% Salt Lake City. Graffity on the walls, most channels are underexposed.
3 low high 1+GPS 2 3 100% Salt Lake City. Exchange Pl., most channels are underexposed.
3 low high 1+GPS 2 3 100% Salt Lake City, Main St., Gallivan Plaza.
3 low high 1+GPS 2 3 100% Salt Lake City at night, Main St./North Temple St., low exposure of HDR pair.
3 low high 1+GPS 2 3 100% Salt Lake City at night, Main St./North Temple St., high exposure of HDR pair.
3 low high 1+GPS 2 3 100% Salt Lake City at night, Main St./South Temple St., low exposure of HDR pair.
3 low high 1+GPS 2 3 100% Salt Lake City at night, Main St./South Temple St., high exposure of HDR pair.

Leave a Reply


2 + = five