Well detection using the Hough transformation is not robust.

[wheelern]

We use the Hough transform to detect circles. It works well for test data, but will not be robust because of the following:

1. Requires careful cropping of the input image
2. Filters circle centers that don't fit within an opinionated grid
3. Variation in the brightness of the well walls.

This the same general approach used in WormAssay, but his is more sophisticated and definitely more robust. Ideas to steal:

• [ ] Use a pre-defined well radius that is adjusted based on plate format
• [ ] Use an errorTolerance to allow the function to search a larger range of radii (i.e. here)
• [ ] Filter by center collinearity rather than an opinionated grid (i.e. here)
• [ ] Switch from scikit-image to OpenCV for the Hough transform

Another idea for filtration is to find more circles than needed, get the distribution, and only keep coordinates that lie within the peaks (number of peaks is the same as number of rows or columns) +/- some error.

[wheelern]

It might take awhile to get this right. An option until then is to only support 24-well plates and provide instructions for cropping.