Look into why dehaze is not working and see if we can accelerate it with OpenCL

[russell-taylor]

[russell-taylor]

I needed to downgrade from numpy version 1.17.1 to 1.16.1 to get dehaze to work.

When I run only that algorithm on the bubbles image with a single processor, it takes 156 seconds. Running with multiple core does not seem to change the CPU load after the first few seconds and it takes just as long.

It looks like it is running a convolutional network across the whole image.

[russell-taylor]

It looks like the code is all running on Numpy arrays and is largely using the scipy.signal library convolve2d function. We cannot turn those into UMats because the convolv2d function expects 2D arrays.