maelstro / io_image_edge_detection_aco Goto Github PK

Implement Ant Colony algorithm in Python, without using Cython/CUDA

For every module, we should have unit tests written to check, if we didn't break anything during the works. It's a hassle, but it's a good practice. For now we won't be setting a minimal code coverage value and no CI/CD, but we may try to set up ie. a CircleCI in the foreseeable future.

I can write a Docker image for testing purposes (tests are run on a VM in most of CI/CD tools, and we have to provide a machine).

As mentioned during the meeting, we should explore other ways of implementing an ACO. Variants proposed:

• Fast Ant algorithm
• No-horizon optimization (already implemented, we don't use the horizon)

If any of us comes across an implementation of one of the mentioned articles, he should post a link to it in the comments below.

When we have a fully implemented ACO algorithm (#2), we're ought to run a grid search to get the best performing set of parameters.

Test image: Lenna

The article mentions that Otsu thresholding method is applied, but during the tests it seems that it gives hardly any output, the threshold is too strict. We should try to develop our own thresholding method, now we have 3 methods:

• Simple threshold: if pheromone value for a given pixel is lower than a tau value, it's classified as the background,
• Otsu thresholding: threshold value is calculated (using skimage.filters.otsu_threshold) and the pheromone value is checked against it,
• According to this code, lines 201-202.
  We should try, which one gives the best results (currently the simple threshold is the best).

Generate groundtruth images for algorithm validation

We need an interface for calculating the edge detection quality.

Requirements:

• Pass an input image and groundtruth image (can be pair of images),
• Return the metrics value
• Log everything to file (logging module in Python)
• CLI-based command
• Allow the user to select a metric or calculate all of them (only if we'll be having multiple metrics)