7. Hessian Data Augmentation - SummerBigData/Iceberg GitHub Wiki

While I was looking for ways to capture the data in the image more effectively, I came across the Hessian filter from scikit. The filter is very easy to implement and appears to peer through the noise and pick out the objects inside the image. Here is an example of what it can do to 18 images in the dataset (with strength 0.05):

Another way to see this is by seperating the channels:

To get a better idea of what the original images looked like, I stacked many hessian filtered images of varying strengths on top of each other to get this:

The goal here is to include this hessian layered image as the third channel to the CNN in hopes of better classification probabilities. Another possibility is to discard the original images entirely and just send the hessian layered image to the CNN. Finally, one could use a similar approach to the one used in the PlantSeedlings challenge; use the weakest hessian filter (large patches) to stamp out the iceberg from the original images to create images with significantly less noise.