Thursday, December 06, 2018

Continuation: Deep Neural Network Search for Better CFA and Demosaicing Algorithm

Thanks to Offline Dreams mentioned another machine learning CFA pattern optimization in the comment to my yesterday's post. "Deep Joint Design of Color Filter Arrays and Demosaicing" paper by Bernardo Henz, Eduardo S. L. Gastal, and Manuel M. Oliveira from Brazilian Instituto de Inform├ítica – UFRGS differs from the previous post in a number of ways:
  • both noisy and noiseless cases are explored
  • CFA pattern is optimized together with demosaicing algorithm
  • different CFA colors were a part of optimization too

"We present a convolutional neural network architecture for performing joint design of color filter array (CFA) patterns and demosaicing. Our generic model allows the training of CFAs of arbitrary sizes, optimizing each color filter over the entire RGB color space. The patterns and algorithms produced by our method provide high-quality color reconstructions. We demonstrate the effectiveness of our approach by showing that its results achieve higher PSNR than the ones obtained with state-of-the-art techniques on all standard demosaicing datasets, both for noise-free and noisy scenarios. Our method can also be used to obtain demosaicing strategies for pre-defined CFAs, such as the Bayer pattern, for which our results also surpass even the demosaicing algorithms specifically designed for such a pattern."

The machine learning optimization picked quite different patterns from the Bayer CFA, both in color and in size:


  1. I wonder if the CFA is designed by machine, is it patentable as a human invention? Who is the inventor in this case?

    1. I think the deep neural network is just a tool here. It's similar to other tools that humans can use while working on inventions: soldering iron, pliers, hummer, etc. So, the invention should belong to a human who operates it.

    2. I don't agree. From what I understand, it was not like the authors had an idea for a new CFA, and used the NN to optimize it. It sounds more like the machine tried many variations across a large design space and came up with an optimization that was not anticipated by the authors. But anyway, this will be a problem for USPTO. I will ask them next time I am down there. I already had a talk with Andrei Iancu about the USPTO's stance about possible future machine generated patents and claims. He agreed it was in a different sort of space from human invention and new territory for USPTO.

    3. OK, I understand what you are saying. Let me try another line of argument.

      When a company asks its employee use his/her biological NN to develop something, the resulting invention is assigned to the company, right? Similar to the company, a human provides to AI all the resources to do the work and sets the targets. So, the invention should be assigned to the human or whatever organization he/she belongs to.

      The next question is who should be named as the inventor. Since AI does not have a name or a ID and is fully dependent on the human who programs and tweaks it, the inventor should be human too, in my opinion.

    4. Well, you could program a name generator too, and solve the name problem. That one is easy. But the deeper philosophical (or legal) issue is if an invention made by a machine is patentable? Patentable is different from assignable.
      So let's consider something else. A company owns a chimpanzee. The chimp is given a puzzle and solves using some new approach. Assuming this meets the general definition of an invention, is the chimp the legal inventor? I don't think animals are legal entities under US law (although they are protected). So a machine or algorithm is even less physical and I am pretty sure they have zero legal status. I am pretty sure that the entirety of patent law, from the beginning of time, addressed humans only, as inventive entities. The rationale for patents, however, remains valid. (the invention is made public to teach others, in exchange for a number of years of exclusive use of the invention)

    5. Hmm, if a grad student invents a new CFA pattern, is that patentable as the supervisor's co-invention?

    6. Seems like you missed the point of the discussion. But anyway, if the person made an inventive contribution reflected in the claims then yes, he/she is a co-inventor. If not, then no. But you should consult an attorney as this is a legal matter and I am not an attorney.

    7. There is jurisdiction in the US which says that animals don't have the legal authority hold copyright claims: the known monkey selfie case. That probably applies to other IP rights like patents too.
      But it seems that monkey took some nice pictures.

      But the question above also brings up the issue of the definition of an 'inventive step'. In EU patent law, a patent can only be issued if there is an inventive step. Now, if a machine is able come up with the solution, was there still an inventive step to come to the invention? And if not, can you then invalidate a patent if an AI machine is able to reproduce the idea? In the US it is non-obviousness, which is considered similar but an AI machine may be able to come up with a non-obvious solution. Difficult questions.

  2. Just wondering how saturated information are managed.
    how to select pattern versus noise? can not change resin on a sensor?
    did I miss something?

    does it mean that random retina light-sensitive cells are not optmized?

  3. in the 4x4 patterns the R+G+B does not add up to 1. Should this mean that a low transmission cfa has an advantage?


All comments are moderated to avoid spam and personal attacks.