1
$\begingroup$

I have an idea but I am not certain that it can be modeled in a DL architecture.

Let's say we have images of different qualities based on color patterns and their assessment as labels in a range from 0-1. E.g. Image 1 has 0.25 quality, Image 2 has 0.5 quality and so on.

Could this be implemented in a standard Resnet50 architecture with 1 output and sigmoid? Is there anything in literature that you could point me up to? I wasn't able to find anything, maybe I am searching it wrong.

EDIT: I have found this https://github.com/idealo/image-quality-assessment, but the implementation is different that I suggest. I know that this can work if I add five different quality classes.

I want to know if I can train this with 1 class output with a rank, in order to let the model understand how those qualities are linked. For instance quality 2 is the next best quality than quality 1.

$\endgroup$
  • $\begingroup$ Could you maybe elaborate what quality is in this case? Are you fi interested in an assessment on a technical level (the amount of noise for instance), functional (can I distinguish what it is) or artistic (Is it interesting)? $\endgroup$ – S van Balen Mar 5 at 11:38
  • $\begingroup$ Ok I see you aded a little more context. How will you obtain these assessments? $\endgroup$ – S van Balen Mar 5 at 11:44
  • 1
    $\begingroup$ Manual annotation on the images. What the user assess as Quality 1, 2 etc. $\endgroup$ – ClosedLoopControl Mar 5 at 11:51
  • $\begingroup$ @SvanBalen Manual annotation. Please check also the second edit. $\endgroup$ – ClosedLoopControl Mar 5 at 11:56
  • $\begingroup$ Is that assessment in anyway guided (with a briefing for instance) or is it more like broad stroked "I like this image"? $\endgroup$ – S van Balen Mar 5 at 13:13
0
$\begingroup$

I found the answer I was looking for in this Google paper:

However, in the case of ordered-classes (e.g. aesthetic and quality estimation), cross-entropy loss lacks the inter-class relationships between score buckets. One might argue that ordered-classes can be represented by a real number, and consequently, can be learned through a regression framework. Yet, it has been shown that for ordered classes, the classification frameworks can outperform regression models [21], [31]. Hou et al. [21] show that training on datasets with intrinsic ordering between classes can benefit from EMDbased losses. These loss functions penalize mis-classifications according to class distances

| improve this answer | |
$\endgroup$

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.