2
$\begingroup$

I'm a newbie in machine learning, so excuse me in advance). I have an idea to make NN that can estimate visual pleasantness of arbitrary image. Like you have a bunch of images that you like, you train NN on them, then you show some random picture to NN and it estimates whether you'll like it or not. I wonder if there is any pervious effort made in this direction.

$\endgroup$
3
$\begingroup$

This question reminds me of a project I saw that used Deep Learning to rate selfies on twitter. But a quick google search shows that there are plenty of projects that are much closer to what you are interested in:

Rating Image Aesthetics using Deep Learning

Predicting Image Aesthetics with Deep Learning

Deep Understanding of Image Aesthetics (with data and model linked)

Understanding Aesthetics with Deep Learning

and probably dozens more.

Of course if you are interested in predicting subjective pleasantness the above is only a beginning. In that case you may also take a look at recommender systems.

$\endgroup$
2
$\begingroup$

I see a main concern with the problem you show and that is the subjectiveness of the term like, what I like is not the same of what you don't like, maybe I like more ciricular shapes and lou like best rectangular ones. The main problem is that with such an subjective label, is difficult to create a global model.

$\endgroup$
2
  • $\begingroup$ Very good comment. I've met people who spend money on work I think is gauche, and by the same token, they find my taste to be bewildering. It seems to me that subjectivity is the value here, in that, such an engine could be used to match people of particular tastes with work that would likely appeal to them. $\endgroup$
    – DukeZhou
    Apr 5 '17 at 16:54
  • 1
    $\begingroup$ Of course, a dataset well designed and appropriate in terms of size, would eventually find, clusters of likeness between different individuals. But, and it is my opinion, I think the problem is far more complex than we can imagine, since there is no absolute truth, and NN struggle with problems that have an absolute truth, like object recognition, so maybe could be one of the challenging problems in the future of Affective Computing $\endgroup$ Apr 5 '17 at 21:54
1
$\begingroup$

That sounds like a pretty straightforward application of a NN classifier to me. I don't know if anybody has done that specific thing or not, but I don't see any particular reason to think it wouldn't work. My advice to you is to just jump in and do it.

$\endgroup$
1
$\begingroup$

I don't think anyone has done it yet,but you could try. A way you could implement it is having a quite efficient CNN trained on the things you like,then your program should ask the user if he does like some images and on the answers he will give, your program will finetune the original network and then with the fresh-trained one you should obtain good results.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.