How the Best Super-Resolution is Different
When deciding on using Super-Resolution in your product, you would normally want to see an objective comparison of different SR solutions. How is one solution better than another? Whilst it is a valid question, making such a comparison is a challenge. Typically, SR is deeply integrated into the ISP pipeline, which means that comparing two solutions would require implementing both in the same device. Imagine, say, a smartphone OEM, which implements two or more SR solutions from different vendors in the same device design. Probably, the only reason for that would be an internal evaluation, results of which never go public.
There could be another way though – when two device designs implementing two different SR solutions share the same camera hardware design. Luckily, such a pair of devices has recently appeared on the market. One is the Qualcomm Smartphone for Snapdragon Insiders, a device by Qualcomm, packed with top-notch solutions for the Snapdragon platform. As the best-in-class solution, Almalence Super-Resolution was selected for that device. The actual hardware design was done by ASUS, who reused the camera hardware from their previous model, Asus Zenfone 7 Pro, powered by an SR solution from another vendor – the biggest camera software vendor on the market.
Surely, there are other differences in the ISP pipeline between those two devices, so there will be a difference in exposure, color, dynamic range, etc. However, the primary image quality characteristics such as details and noise can be objectively compared.
Let’s take a look at a few examples and see how the Super Super-Resolution differs from the other methods.
Cash bills are always a nice target for SR algorithms comparison, lots of fine details:
The text also makes a good target, for us to be able to see how readable it is, and also to distinguish real details from noise looking as "details". In this example, we see an attempt to "boost" the performance of a weak Super-Resolution solution by oversharpening, so that the image should "look sharper". No, it does not work. Sharpening is not an effective resolution increase - it does not capture fine details, and just makes the bigger details look worse:
Super-Resolution is not only about reconstructing the details. A big part of an SR solution is improving the detail/noise tradeoff. In other words, filtering the noise without losing the fine details. A weaker SR method requires stronger noise post-filtering, which washes out details. As the color noise is most visible and annoying, it requires stronger filtering, thus the loss of color detail is drastic. A real Super-Resolution avoids it by reconstructing more details in the first place and better filtering the noise then:
Dea Leaves aka Spilled Coins chart prevents confusing sharpening for resolution increase:
Get in touch with us to find out how Almalence Super-Resolution can differentiate your imaging product by squeezing the most of your camera hardware and achieving the best resolution and SNR!