How the model works
The image classifier behind ChickDetect: how I trained it and how it performs on photos it never saw.
-
Data
10,467 photos in 5 classes: four kinds of dropping plus “not a dropping”, so the app can say when a photo is something else. Split 70 / 15 / 15, with copies of the same photo kept in the same split so the test score isn't inflated.
-
Model
Transfer learning on MobileNetV2, pretrained on ImageNet and used as a frozen feature extractor. I trained a small classification head with dropout, class weights and augmented images (flips, rotations, darker exposures).
-
Ship
One 10 MB TensorFlow Lite file that takes a raw photo and returns five probabilities. Pixel scaling lives inside the model, so the app and the model can't disagree. It runs offline on the phone.
-
Test
Scored as the exported file, the same one the app runs, on 1,572 held-out test photos.