Canny is definitely the best. Others will be only worse. To cope with fluctuations, use neural nets (though, not in OpenCV yet).
@ya_ocv_user, There is one network called HED (holistically-nested edge detection network) which represented in the turorial https://docs.opencv.org/master/dc/db1... (a sample script is https://github.com/opencv/opencv/blob...).
Deep learning again ... You are really zombied. So what should user do? Spend a week and read how to add a layer to DNN and how to use various DL products, or just write a dozen lines of code and implement this layer alone? What you recommended is a scientific publication and you perfectly know that 90% of these publications are practically useless.
I don't think that it is a good idea to present all neural nets under the brand "Deep learning". NN are a completely different computing style (in comparison to traditional programming). Yes, they may be trained, may be used for concept formation, but there are solutions which don't require learning. By the way, this is true for human vision too. Everything is defined by synaptic connections. In humans, they depend on 2 factors - individual experience and inborn, genetically predetermined features. What is known as convolution, that is the most useful short-distance links, belong mainly to 2 category.
Asked: 2018-06-01 00:06:48 -0600
Seen: 1,728 times
Last updated: Jun 01 '18
You can use HSV color scale to remove lighting sensetivity