Open svepe opened 9 years ago
The images are shadowgraphs so brightness is constant I think, right?
Finding orientation is non-trivial as plenty of the classes have radial symmetry.
It would be nice to keep track of the original organism pixel size before resizing so we can make the most of the constant pixel/actual length ratio in the dataset not sure how this would be incorporated into CNN (maybe just added to the matrix and padded).
What I plan to do is fit an ellipse to the plankton, rotate the image such that the long radius points north and resize such that the image becomes circle. The radii of the ellipse could be used as size input in some of the higher layers.
Find the size, position and orientation of the plankton in each image and rotate and resize the image accordingly. Normalise the brightness of each image as well.