Wednesday, December 10, 2008

Chipworks Reveals Omnivision Split Pixel Internals

Chipworks did a first round of analysis on Omnivision split pixel sensor back in April and kindly agreed to shed some light on the subject even though they haven't gone forward with a full analysis. Chipworks has posted a flyer based on the brief analysis, a custom landing page can be found here. (This page is temporary, so keep the pdf file, if you need it.)

The design is quite interesting. Each pixel has two photodiodes per each 6um pixel, as depicted below:

Each split pixel has dual microlens above, one per each photodiode (I think Visera spent many sleepless nights to optimize these oval lens shapes):

It looks to me that Omnivision used dual exposure technique to achieve its goal of 110db DR with this sensor. Two interleaved rolling shutters can run across the array, where reset curtain is followed by two read curtains - short and long one. Then the two readouts are combined into a single HDR image. Obviously, there should a large memory to hold the intermediate frame results. The die photograph in Chipworks document supports this guess, showing a huge memory-dominated digital portion.

So, to me this looks like a dual exposure sensor with on-chip HDR reconstruction. Its main competition are Pixim digital pixel and Toshiba-like dual exposure pixels with charge skimming.

In principle Pixim can get better dynamic range, but I believe Omnivision 4T pixel is better in low light, especially if two photodiodes operate in normal mode, with same exposure on the two photodiodes.

Comparing it with Toshiba solution, I'm less sure about Omnivision advantages. Toshiba sensor might have some linearity artifacts when connecting short and long exposure ranges. Omnivision probably does not have this. However, Toshiba has an advantage in low-light sensitivity in HDR mode because Omnivision sensor effectively loses half of the light for the short exposured photodiode.

One question that bothers me though, is that the two photodiodes are spacially shifted relative each other. So, there might be some color artifacts when interpolating the edges on HDR picture. It's probably not something fundamental, but might take some effort to correct.

Many thanks to R.F. from Chipworks for providing this presentation!


  1. I don't think you need any frame storage with this approach. You'd just run the reset curtains staggered with respect to each other and keep the read curtains aligned so that the read of each pair of pixels occurs during the same row time. You get 2 different exposures for a given bayer tile that can be combined in a LSBS fashion inline. They probably need a few row buffers and a lot of digital logic to make it work, but the amount of memory is not excessive. This technique is efficient in that the exposures can occur simultaneously, but dark pixels may loose nearly half their photons in some scenarios. This is the cost of doing it inline. - KF

  2. Yes, indeed, having read curtains aligned saves the memory. However, it makes blooming problems worse.
    If reset curtains are aligned, the short exposed photodiode is read first, so any blooming from the long-exposed one does not matter.
    If read curtains are aligned, blooming from the long exposed photodiode might bloom to the short exposed one. So one needs to deal with it.


All comments are moderated to avoid spam and personal attacks.