The Minds Behind the Magic Special Interview DIGIC 4

part 1, Parallel processing technology - the key to DIGIC 4’s enhanced processing speed

The evolution of the DIGIC image processor

Who has been involved in the design of the DIGIC digital camera image processor?
Ikeda
There are several teams engaged in the development of new image-processing technologies and new image-capture functions. For example, there had been people working on the face detection function for some time, and because this function was included in DIGIC III, we were able to build on this and further improve this function for DIGIC 4.
Besides cutting-edge research on our key technologies, there were other teams who were always working on peripheral technologies that could possibly benefit DIGIC. Then there was Sekiguchi's team, which works on system-in-package chip design and circuit board layout, and other teams that work on individual designs for incorporating DIGIC into actual products.
I understand that the effort involved several different teams, but what, exactly, was the development theme for the DIGIC 4?
Ikeda
Our goal was to achieve further improvements in image quality and processing speed, upgrade and strengthen the functions that DIGIC III provided, and also add new functions. Put simply, achieving higher image quality is a battle against noise. If you increase the sensor pixel count, then with low sensitivity you get a high-resolution image. However, if you set a high-sensitivity compact camera to ISO 1600, the signal from the sensor will inevitably be full of noise. The key development challenge as far as image quality is concerned is eliminating that noise from the signal!
So you were focusing on noise reduction. Could you be more specific?
Ikeda
There are a wide range of algorithms that can be used to achieve noise reduction. By putting these algorithms together in complex combinations, it is actually possible to reduce noise very effectively. By the time that DIGIC 4 was introduced, I believe that we had achieved the best possible results that we could have hoped for in terms of noise reduction.
But surely more complex algorithms require more processing time...
Ikeda
Yes, that's right. If you used the DIGIC 4 algorithms with the speed and chip size of DIGIC III, it would probably take as long as 3 to 4 seconds to process a single image, or maybe even as long as 10 seconds. Because that wouldn't be commercially viable, it was vitally important to improve processing speed. To put it another way, it was only because we succeeded in improving processing speed that we were able to improve image quality as well.
The LSI chip area of the DIGIC 4 processor is larger than with DIGIC III. Is that why you were able to improve the processing speed?
Rengakuji
It's true that the area available for circuits on the LSI has been increased, and the scale of the circuits and the scale of the algorithms have more than doubled. However, that isn't the main factor behind the improved processing speed. The reason why we were able to achieve our processing speed target was because of new circuit design technology.
It was Kawaguchi's team that set the processing speed target for DIGIC 4, wasn't it?
Kawaguchi
Er, yes, that's right (laughs). The image quality and processing speed requirements were very demanding. To meet them, we had to put a lot of effort into optimizing the circuit design. We decided to use a parallel processing circuit design, which would provide smoother processing flow without eliminating any of the processing circuits needed for high image quality. We used parallel processing not only for the auto-focus, auto-exposure and LCD display functions, but also for face detection. By adjusting the algorithms, DIGIC 4's face detection function was made to work three times faster when compared with DIGIC III. By adopting this design technology, we were able to achieve high-speed processing of the algorithms for improved image quality that Ikeda asked my team to develop.

Sekiguchi

Sekiguchi
If you use parallel processing, there is an increase in data flow between the LSI and the memory outside the LSI. As the team responsible for DIGIC 4's system-in-package design, this gave us a lot of headaches (laughs)! By employing a range of different package and circuit board technologies, we were able to achieve an improvement in signal integrity when transmitting data.
So far we've only been talking about hardware circuits. Is software not important for image processing?
Ikeda
What the DIGIC software handles is command control, with the main focus being on "directing traffic." The software has nothing to do with the actual image processing. If you used software for image processing, it would take too long. For example, if you tried to process a single image from a 10 MP sensor using an ordinary desktop PC, it would take about fifteen minutes to complete. DIGIC 4's hardware circuits do it almost instantaneously. We've reached the stage where you can perform continuous shooting at 10 frames per second.
Really? So hardware circuit processing is that fast!
Canon Technology

Top of Page

  • News Feed