Unsupervised ML for detection of ruler unit markers and pixel-to-metric conversion factors

Most archival specimen images include a scale bar or ruler. Detecting rulers and cropping them out of the image is straightforward, but accurately detecting unit markers for pixel-to-metric distance conversion is more challenging than it appears. LeafMachine2 has already pushed heuristics and traditional computer vision methods to their limits. Now it’s time to explore a more robust and flexible approach: detecting periodic unit markers with unsupervised ML methods.

We have two datasets of thousands of rulers that vary in appearance, condition, resolution, and units:
[Cropped Rulers - https://huggingface.co/datasets/phyloforfun/LM2rulers38](https://huggingface.co/datasets/phyloforfun/LM2rulers38)
['Squarified' Rulers for Class Detection - https://huggingface.co/datasets/phyloforfun/LM2rulers38squarified](https://huggingface.co/datasets/phyloforfun/LM2rulers38squarified)

The challenge is to design a custom autoencoder (or similar unsupervised model) that can detect tick marks without requiring labeled datasets. Once the ticks are identified, we can measure periodicity and calculate how many pixels equal 1 cm in any specimen image.

This project could lead to the development of a much-needed tool that would significantly improve our ability to perform large-scale quantitative trait extraction.

Start by inspecting the Cropped Rulers and 'Squarified' Rulers datasets linked in the issue and the repository structure, since no implementation files or tests are named. Explore an unsupervised model such as a custom autoencoder for periodic unit-marker detection. Done means detecting ruler ticks without labeled data and using their periodicity to calculate pixels per centimeter in specimen images.

Written by the indexing model from the issue text.