On differentiability of common image processing algorithms
Alexander E. Zhukovsky · 2019
We present differentiable implementations of several common image processing algorithms: Canny edge detector, Niblack thresholding and Harris corner detector. The implementations are presented in the form of fully convolutional networks and explicitly arranged exactly to the original algorithms. Usage of such form of the algorithms allows to tune their parameters with a gradient descent. We performed parameter tuning in the edge detection problem and it shows that our implementation enables us to obtain better results on the BSDS-500 dataset. As a part of implementations of algorithms, we introduce a generalization of pooling method, which allows using arbitrary structure element. We also analyze the given neural network architectures and show the connections with contemporary approaches.