1
$\begingroup$

I wrote a script that detects objects, crops the part of the image where an object is located and tries to increase the quality of cropped image. The object detection runs tensorflow. To increase image quality I use the Google RAISR AI. It tries to increase the image's resolution by assuming "in-between-pixels" through its algorithm. Although this method is not perfect, it fits my need. After that, a threshold of the image is built for further analyzing.

Original image with detected object (upper line in object = reflection):

enter image description here

Cropped object image:

enter image description here

Improved image with RAISR:

enter image description here

Threshold:

enter image description here

With these images, my script can create a threshold. The main aspect of success is the quality of the object as well as the quality of its reflection on the inner electrode of the pipe.

Now I have datasets which show a much weaker reflection of the object, for example when the object is not at the inner electrode but at the outer wall.

Original image with detected object:

enter image description here

Cropped object image:

enter image description here

Improved image with RAISR:

enter image description here

As you can see, no reflection is visible on this image with the human eye.

This makes it difficult to build a threshold:

enter image description here

The reflections can only be seen on a few images of the dataset. These could be used to build a threshold. Unfortunately 98% of the images are not of this "quality".

enter image description here enter image description here enter image description here

I am looking for a method to "harvest" as much information as possible from images like these:

enter image description here

and build a threshold. Are there methods to make the script "see" the reflection of the object although it can't be seen with the human eye? This image should result in this threshold:

enter image description here

I know that my wording of the problem might be a bit unprofessional. I hope I could make it clear.

$\endgroup$
  • $\begingroup$ I would say that the title is a bit misleading here, because the objective is to segment or detect the shard...not "increase" any amount of information found in the image. By the way, object detection can be done with a variation of the networks that are used to detect the shard. If you go back one level, the network is probably creating an aggregation of pixel patches that are not pipe (most probably). So, it is likely that you can extract a better outline of the shard by tracking those regions. Can the shards be found in any place in the pipe?(e.g. top, hanging from the left, right, etc) $\endgroup$ – A_A May 2 at 11:50

Your Answer

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

Browse other questions tagged or ask your own question.