Dave Etchells over at
Imaging Resource has published an interview with a group of Fujifilm managers and engineers regarding the technology in their latest X-Trans cameras.
Machine-learning is increasingly working its way into each new generation of digital cameras (albeit not quite
to this extent yet) and I thought the description of what is involved in training a neural network to perform various forms of subject detection was particularly interesting:
RDE [i.e., Etchells]: I’m especially interested in the deep learning aspects of the AF system. Those would take a long time to develop, so I’m curious how long ago you started building the data set and training the algorithm.
Jun Watanabe: The number of photos used in the training set was more than 10,000 photos. The process of AI development is that first we separated the different parts of each image, for example the eyes, the face, the head and so on.
RDE: So a human had to go through every image first, saying “this is the head, this an eye”, etc.
Jun Watanabe: Yes. The second part is to use different sizes of the images and different angles, so it can detect the subject more accurately. Then the last thing was that we did field tests, to see if there were any subjects that were hard to detect, and we’d do more training for those subjects.
RDE: It must be a very long process, I think - one year, two years, three years?
Makoto Oishi: I think just after the development of the last processor, we already started to investigate the current device so I think at that point we already decided to incorporate the AI processor in the new chip.
RDE: So basically, as soon as the previous processor was shipping, then you were thinking about what would come next and had some idea that it would involve AI and deep learning.
Yuji Igarishi: I think the number of images wasn’t just 10,000 images, but tens of thousands. . . .
RDE: It’s interesting, I read that the bird AF also works very well for things like frogs, dragonflies, butterflies, etc. Did you train on any other kinds of subjects besides birds,, or was that just a happy accident?
Yuji Igarishi: Happy accident, coincidence.
While each generation of on-board processing technology offers more opportunities for cameras to exploit "artificial intelligence," I presume the manufacturers inevitably will run into limitations based on power consumption. Of course, where the capture device leaves off, the post-processing software often can take over. Given recent increases in affordable processing and graphics cycles, I suspect we still have a long way to go before the general-purpose computers we use in post begin to hit a wall.
