Granny – image based analysis of fruit quality; TR-24-100A
Author: Loren Honaas
Published: 2026
Summary: This project delivered an important advancement in automated image-based fruit quality assessment software. Granny, an open-source platform originally developed to rate superficial scald, has now been fully re-engineered into a modern, modular system capable of robust and reproducible apple starch clearing evaluation. Across the two-year period, the project team completed a full rewrite of the pre-existing disjoint software into a single unified Python package, implemented a modular expansible software framework, and integrated a modern deep-learning model for identification and segmentation of individual fruit on a tray. The primary objective of this project was to deliver a starch-rating module that provides granular and consistent ratings, avoiding the pitfalls of human rater bias. Additional rating modules for superficial scald, pear color, and pear blush were also upgraded and integrated. Extensive functional testing ensures reliability across future versions. Image-based starch ratings produced by Granny correlate strongly with human rater assessments across commercial samples of ‘Envy’ and ‘Honeycrisp’ fruit. Its design allows industry users and researchers to expand its capabilities by integrating custom analysis modules and database systems. Industry collaboration was central to success. Growers and packers provided images for training and validation, and several partners tested the software in their internal workflows. User documentation and developer guides were substantially enhanced, and design support was provided for imaging stations and tray-layout templates. A new graphical web interface—supported by a supplemental USDA award and developed in collaboration with the University of Tennessee—extends access to Windows, macOS, and field-use environments without requiring installation. This work provides to researchers and the industry, objective, standardized, and scalable fruit maturity measures and provides a flexible platform for additional trait-rating modules.
