Vein segmentation of pre-segmented grapevine leaves using the radial Euler Characteristic Transform

For grapevine leaves pre-segmented from a background, code for a UNet model to segment venation using the Euler Characteristic Transform (ECT). Training data is derived from the `data` folder found in the following repository, and is included in this download: https://zenodo.org/records/17143420. Blade and vein outlines are used to extract leaves and create a pre-segmented leaf on a background in a 512x512 png image for training. From the extracted contour of the leaf, an aligned radial ECT is created as a second channel to predict the vein mask. The corresponding vein ECT is also used as an auxillary during training for predicting the vein mask. Inference is intended for nearly 25,000 pre-segmented grapevine leaves stored as 512x512 RGBA images segmented in the alpha channel. These images for inference are not included in this download, and the folder `FINAL_MASKS` can be downloaded and placed in the project folder from here: https://zenodo.org/records/16883403.

The following is a summary showing the relationship of this model to others:

Segmentation of RGB leaf masks

• LEAF_SEGMENTATION: https://zenodo.org/records/16883916
• FINAL_MASKS: https://zenodo.org/records/16883403

Segmentation of full vein mask

• Vein segmentation of pre-segmented grapevine leaves using the radial Euler Characteristic Transform: https://zenodo.org/records/17741713

Segmentation of primary veins

• Primary vein segmentation in grapevine leaves: https://zenodo.org/records/17881166

Segmentation of midvein

• Midvein segmentation of grapevine leaves: https://zenodo.org/records/18005383

Gaussian heatmap estimation of petiolar junction coordinate

• Grapevine leaf petiolar junction detection: https://zenodo.org/records/17922822

Segmentation of midvein, distal, and proximal lobes

• Grapevine leaf lobe detection model: https://zenodo.org/records/18004938