Natural images are often associated with textures. From most edge detectors I have come across, the emphasis is usually not on rejecting textures. For complex images, both textures and boundaries are captured. I wanted to have an algorithm (or an idea) that only detects boundaries of objects, like in a hand-drawn art drawing. Any one with idea, please share.
For sure this can not be achieved by a low-level edge detector. You have to introduce high-level semantics about the scene into your method. The specific computer vision field you are aiming at is probably object boundary detection. I would suggest the read this paper: http://www.cv-foundation.org/openaccess/content_cvpr_2015/papers/Uijlings_Situational_Object_Boundary_2015_CVPR_paper.pdf
You could involve machine learning. A neural net could
have a look at a point on an edge and the environment
of it. Instead of a neural net a different ML algorithm
might also suffice, e.g. comparing histograms against
pairs of histograms and their classification in a list.
While the view focus of the neural net wanders
along a suggested edge, it could continuously adapt
its notion of the edge. No offline-learning would
be necessary, the ML algorithms would train
itself and learn during it processes one image
and then could possibly forget.
Regards,
Joachim
you can also used the texture based segmentation technique based on entropy. It will give both the positive and negative image. You can easily extract the foreground. I think it works
From time to time, we were coming to this problem in our team. The question You asked can be, in fact, reformulated as how to detect textures and how to discriminate them from edges (and fine details that are also often detected as potential edges). You can look at some our papers concerning the topic at http://asp.eurasipjournals.springeropen.com/articles/10.1155/ASP.2005.1185 or http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=7357321&url=http%3A%2F%2Fieeexplore.ieee.org%2Fiel7%2F7347238%2F7357246%2F07357321.pdf%3Farnumber%3D7357321
I don't think that there is a very good existing solution although ideas proposed above can be fruitful. The main feature in recognizing standard edges and textures is that much more compactly placed pixels are detected as potentially belonging to edge for textural regions. Probably, this property should be exploited somehow.
something upto date on the learning side as @Joachim Pimiskern suggested.
PUSHING THE BOUNDARIES OF BOUNDARY DETECTION
USING DEEP LEARNING
Iasonas Kokkinos
http://arxiv.org/pdf/1511.07386v2.pdf
another recent one:
Object Contour Detection with a Fully Convolutional Encoder-Decoder Network
http://arxiv.org/pdf/1603.04530.pdf
If the texture is regular it is not so difficult problem to construct specific texture removing filter using characteristics of texture pattern.
In other case there is no common solution and you have to look for a feature/more features, distinguishing texture pattern from edges. This is not trivial problem.
- Badges
- Science topic
- Similar topics
- Mathematics
- Algebra
- Matrix