Suppose (only suppose) I have implemented the following 3rd-party research article for my MSc thesis:

Deep learning and image processing for automated crack detection and defect measurement in underground structures
F. Panella 1,2, *, J. Boehm 1, Y. Loo 2, A. Kaushik 2, D. Gonzalez 2

I have validated their theory and earned some skills in the area (say, tools, programming languages, and so on). And, now, I want to write my own research-paper in order to publish in a peer-reviewed journal.

How can I get started?

2 Answers 2


In general, reproductions of existing research is not publishable. If you want to publish, you usually need to do something new. There are two major ways that you can do new work based on existing work.

  1. Make it better: If the existing approach works very well, you can develop an extension of the methodology that is more general than the original work. Maybe you can make a theoretical improvement, like removing assumptions about the loss function. Maybe you can find new areas where you can apply this methodology to solve other problems it wasn't meant for. Maybe you can make it more accurate, or make it's failure less likely.
  2. Discover it's limitations: If it is a methodology that has limitations that are not widely known, that can be a research contribution too. For example, maybe this crack detection algorithm only works with certain ambient brightness. If it's too dark, it looses the ability to detect cracks. Maybe it makes some assumptions about the shape of cracks that are unfounded, and cracks that are a circle, instead of having distinct end-points, cannot be detected by this methodology.

In both cases, you need to do something that goes beyond what is contained in the paper and do something new. But it's not enough to just do something new, you have to do something that both new and interesting.

"Interesting" often means "in opposition of people's default assumptions." Showing that a cave crack detection algorithm doesn't detect cracks in hardwood floors isn't very interesting if no one expected it to. In that case, showing that it does detect cracks in hardwood floors would likely be interesting and open up many new applications for the approach.

Another way that your work could be interesting is that it solves a problem that people have talked about being interested in already. Maybe it's well-known that the algorithm breaks when it gets very dim, and there are multiple papers saying "it would be nice if we knew how to make this work for dim caves." Even if this hasn't been explicitly stated, if people say "I wish we knew how to detect cracks in dim caves" then showing that this methodology does that would be interesting, assuming the original paper didn't and that figuring out that it does work in this context is non-trivial.

These are not the only two ways to extend word, nor is it the only way to show that your work is interesting. But thinking about the problem in these terms will be a good way for you to get started.

Also, if you wrote a MSc thesis, presumably you have a thesis adviser. That person would be a great resource to talk to, both to learn about what is publishable and to get advice about how to get started doing publishable work.

  • 1
    Is it possible that this was the first verification of the theory and wasn't present in the earlier work?
    – Buffy
    Aug 31, 2018 at 11:15
  • @Buffy Yes, this is necessarily an incomplete list of how the work could become publishable. I think that in general this describes good approaches to get the OP started. If there are other circumstances or approaches that you think are noteworthy enough to mention, you’re welcome to write them in an answer. Aug 31, 2018 at 12:12

In addition to the very good answer of @Stella Bidermann:

  1. Apply it to a different area / change the underlying question a bit.

E.g. "underground structures in the paper means concrete pipes or other human-build structures, you could try it for caves or oil pipelines.

Since you used deep learning, the learning base / input data might also be different and lead to different results - this could be examined as well.

  • change the underlying question a bit. --- rephrase plz.
    – user84565
    Sep 24, 2018 at 4:43

You must log in to answer this question.