Conversion of black & white photographs to colour...

[SeriousSam]

I won't even pretend to understand half of what is discussed in this computing science paper, but the end results are rather impressive.

http://hi.cs.waseda.ac.jp/~iizuka/pr...on_sig2016.pdf

[g8ina]

"Our model consists of four main components: a low-level features network, a mid-level features network, a global features network, and a colorization network. Conceptually, these networks function as follows: First, a common set of shared low-level features are extracted from the image. Using these features, a set of global image features and mid-level image features are computed. Then, the midlevel and the global features are both fused by our proposed “fusion layer” and used as the input to a colorization network that outputs the final chromimance map. Needless to say, this is not explicitly implemented as a sequential procedure; rather, it is realized as a single network. Note that no pre-processing nor post-processing is done: it is all computed in a single step. Additionally, as a side product of our approach, we can also perform classification of the scene. While the global features are computed using fixed-sized images, our novel approach for fusing the global and local features allows our model to be run on input images of arbitrary resolutions, unlike most Convolutional Neural Networks."

Eh ??

[Saracen]

Last time I read someting like that, it was a "simple" explanation of why matter can exist in two places at once, using Quantum Mechanics. Though, a number of bits of EU legislation come pretty close. I wonder if the author has a Ph.D in Befuddlement and Obfuscation, or if they're just naturally a public communications idiot. Either way, may I suggest they apply for a post as Donald's Trump's speechwriter. He could do with a heathly dose of that, to improve his PR.

TL : DR version = ... huh?

[scaryjim]

...

Eh ??

Scarily, I get most of the concepts in that little passage your repeated. No idea how you go about doing it, mind you, but conceptually I see where they're coming from.

... I wonder if the author has a Ph.D in Befuddlement and Obfuscation, or if they're just naturally a public communications idiot. ....

If you have a PhD in just about anything - but particularly if it's in hard sciences - the second part seems to come as standard nowadays. A number of years ago I had several conversations with Diana Garnham of the Science Council in which she bemoaned the steady decline in communication skills in general, and writing skills in particular, amongst scientists over the preceding couple of decades. There's a growing niche for science and research communicators: non-specialists who can grasp just enough of a complex topic to explain it to the public in a meaningful way. Personally I'd rather have scientists who can communicate, but perhaps I'm just old-fashioned...

[Saracen]

I thought it was deliberate policy among scientists? After all, "trades" have been doing it for hundreds of years. It's where Guilds originated, innit?

I thought the logic was to use incomprehensible gobbledygook as a barrier to entry, preserving the mysticism of a "trade" and to convince the peasants that as you could speak with confidence yet be utterly incoherent to non-cognoscenti, that you absolutely must be an expert?

It's certainly been a common trait from the print Guilds trying to keep plebs out with everything from jargon to Copyright law, and today's computerese do much the same with Acronyms and big words. I suspect you'd find the great, great ...... great ancestors, in concept at least, in Egypt several thousand years ago controlling the building of pyramids.

[Zak33]

There's a growing niche for science and research communicators: non-specialists who can grasp just enough of a complex topic to explain it to the public in a meaningful way.

Bill Bryson IS that man...and very wealthy he is too

[CAT-THE-FIFTH]

The research paper is targetted towards others in the field and will be understood by the people it is targetted to. It requires a reasonably in-depth knowledge of the field and the prior research done - some of which is referenced in the paper. The maths is reasonably complex unless you have a background in it.

It is not a general piece for people outside the field or the general public.

"Our model consists of four main components: a low-level features network, a mid-level features network, a global features network, and a colorization network. Conceptually, these networks function as follows: First, a common set of shared low-level features are extracted from the image. Using these features, a set of global image features and mid-level image features are computed. Then, the midlevel and the global features are both fused by our proposed “fusion layer” and used as the input to a colorization network that outputs the final chromimance map. Needless to say, this is not explicitly implemented as a sequential procedure; rather, it is realized as a single network. Note that no pre-processing nor post-processing is done: it is all computed in a single step. Additionally, as a side product of our approach, we can also perform classification of the scene. While the global features are computed using fixed-sized images, our novel approach for fusing the global and local features allows our model to be run on input images of arbitrary resolutions, unlike most Convolutional Neural Networks."

Eh ??

I am not a computer scientist,but still I try and explain it in a simple way(and yes computer scientists I will be generalising things too and might not be using the right terminology).

Right,so the whole point of the software is to automate the process of colourisation of black and white pictures. This basically needs the software to be able to determine what kind of pictures are being scanned and to determine how the colourisation pattern is applied.

So for instance,if you scanned a picture of a Moose - you would expect the software not to colour him a bright,but fetching,pink!

What the researchers are using are these:

https://en.wikipedia.org/wiki/Artificial_neural_network

So basically to do with machine learning. To put it simply they are pieces of software which are meant to be inspired by the way neural nets work in organisms(think of your brain and think of the equivalent in simpler organisms). So,instead of having to programme every input with a specific response,the network attempts to "learn" certain behaviour and this is done by "training" the network.

Training the network is done by using test data which we know will have a specific output,so it knows what is "wrong" and what is "right".

It's like when you are being taught driving - the instructor is essentially trying to steer you alone the correct path.

Okay.

Next step - the image needs to be processed so the correct colourisation pattern can be applied,so the algorithm needs to determine basic features from the image to do this. Once these are identified it can be used to generate more generalised feature maps(basically data points which group the images into certain categories,which the net has encountered before) in the images and these will be actually applied to the colourisation process,to determine what colours are applied to the pictures and in what pattern.

Now,what they are saying is there are three feature determining steps and one colourisation step,and these are all done by one network,not multiple ones in sequence.

The authors point out unlike other prior approaches it is more flexible on image size. They also use this information to be able to classify the image too,ie,could be a landscape one,one with people,one with a galloping Moose,etc.

Hence by "training" this network with more and more images,they can actually increase its accuracy.

Might be not 100% correct at a CS level,but that is what I am getting from it.

If you have a PhD in just about anything - but particularly if it's in hard sciences - the second part seems to come as standard nowadays.

But that is not really fair either - some of the stuff is quite complex and the problem is with explaining more and more difficult subjects to people who have no clue about the area,is you need to generalise. The problem is that those generalisations can be technically incorrect even if they get the gist of the meaning across.

My previous attempt at an explanation is probably filled with errors lol.

But at the same time cross discipline is even harder....

I mean trying to explain Biology to computer scientists was bloody harder than explaining it to people who have zero knowledge of biology. I luckily managed to get it accross LOL.