CEVA's Super-Resolution algorithm enables the creation of high resolution images using low resolution image sensors and allows high-quality digital zoom in real-time on mobile devices. Traditionally, such applications were only available on PC systems, limited to offline processing of pre-captured images.
Eran Briman , VP marketing at CEVA says: "Our new Super-Resolution algorithm for the CEVA-MM3101 platform marks the first time that this technology is available in software for embedded applications. It is a testament to both the expertise of our highly skilled software engineers and to the low power capabilities of our CEVA-MM3101 platform, which comprises the hardware platform together with optimized algorithms, software components, kernel libraries, software multimedia framework and a complete development environment. We continue to lead the industry in the embedded imaging and vision domain and the addition of this latest high performance software component to our platform furthers illustrates the strength of our IP portfolio for advanced multimedia applications."
Click to expand |
Click to expand |
Click to expand. Somehow does not show a significant improvement in resolution. May be the 4 component images were shot from tripod by mistake? |
Update: Once we are at HDR and Super-Resolution algorithms, Chalmers University of Technology, Sweden published a fresh thesis "Towards Joint Super-Resolution and High Dynamic Range Image Reconstruction" by Tomas Bengtsson.
"Super-resolution" or "super-sharpness" ?
ReplyDeleteIn the given examples, I do not see improvements in resolution! Only the sharpness is enhanced, but this is not super-resolution and this can be easily done with a single image.
Not entirely correct. Judging by the resolution chart sample, the linear resolution improved by 5-10%, but it does not look like a 5MP to 20MP jump.
DeleteI'd guess, in this test they put the camera on tripod by mistake, while super-resolution needs a slight shift between the frames, such as one coming from handshake.
Meant to me handheld camera shake, not handshake.
DeleteThese images can't be for real. With the text I can get the exact same image as the Super Res by running a sharpen. The super res image is also full of sharpening halo's. The only benefit I see is less noise after sharpening but no additional resolution.
ReplyDeleteI agree. The image with the text shows significant overshoot artifacts, like what you would expect from USM or even RL-deconvolution sharpening.
ReplyDeleteI have implemented a poor-man's SR algorithm involving image coregistration, followed by stacking at a higher resolution. What I saw was not an increase in resolution, but a reduction in aliasing artifacts. The reduction in aliasing artifacts allowed me to use more agressive sharpening, producing an image that appeared to be sharper, but objective tests (MTF curves on slanted edge targets) showed little or no increase in resolution (depending on how agressive the sharpening was).
But I am no expert on the topic. Other SR algorithms may indeed be able to increase effective resolution.
I fail to understand: are these two different interpolation techniques -- two methods to extrapolate from a set of "low" resolution images or are we seeing one image vs 4 combined images?
ReplyDeleteIn the resolution chart picture, the CEVA Super Res side seems to have a couple of 'block' artifacts in the horizontal pattern.
ReplyDeleteI agree that the images have been 'over sharpened' definite ring on the edges.
In addition, something funky is going on in the vertical '16' bar. It is as if part of the bar has been shifted by 1/4 pixel or so in the middle. (Also evident in the '22' bar.)
The enhancements also don't look the same, there is color fringing in the normal image that is not there in the Super Res image.
If you zoom into the '16' line, you can see some awful blocking artifacts (ringing?) in the gray background.
(I personally like 'softer' images anyway.)
Embedded Super Resolution is already available in mobile devices:
ReplyDeletehttp://blog.almalence.com/superb-quality-zoom-with-newly-available-huawei-ascend-p2-first-tests-at-mwc-2013/
CEVA’s Super-Resolution is focused on resolution improvement. It was compared with several leading PC Super-Resolution applications and showed better line separation quality. Also it shows significant improvement vs. sharpness only. It should be emphasized that CEVA's SR solution is targeting embedded solutions (vs. PC, off-line) with focus on mobile devices, and running fully in SW (meaning C code) in real-time and already optimized to high performance and very low power consumption.
ReplyDeleteRegarding the different comments re-overshoot – as CEVA’s SR is SW-based, it supports various configuration options which can adapt such parameters to the user’s preferences.
In the link below CEVA’s SR is compared to a leading PC SR application, using a resolution chart image taken by a mobile device.
http://www.ceva-dsp.com/images/stories/Diagrams/compare_ceva_sr_vs_pc_sr2.png
Granted, the CEVA is sharper, but the comments above still stand. There are funky blocking artifacts, ringing, and something was done to the dynamic range between the samples. The background of the CEVA image is 'darker' and the blacks are darker than the standard image.
DeleteAgain, it might be just me, but I don't see how the link shows anything different from the previous images.
Yes indeed, their new image is darker, so not an apple-to-apple comparison, but to me it looks much more crisp and no artifacts (or very few).
DeleteMaybe it's just me...
Well, relating to the image being darker, we should compare to the original source image. seems the CEVA color is the very close to original, unlie the PC SR app. see
Deletehttp://www.ceva-dsp.com/images/stories/Diagrams/compare_res_chart_color_ceva_vs_pc_sr.png
Yair, how did you make the comparison with 'Leading PC application'?
DeleteThat application requires super resolution profile for particular camera. As far as I know, there is no profiles for RAW images for mobile devices in that application. With no profile the comparison has no sense.