All work

Case study

ChickDetect

An Android app that classifies photos of poultry droppings as healthy or showing signs of coccidiosis, Salmonella or Newcastle disease. I trained the image model behind it, and it runs offline on the phone.

  • Flutter
  • TensorFlow Lite
  • MobileNetV2
  • Python
A hand holding a phone running ChickDetect over a chicken dropping, with a chicken in the background

How the model works

The image classifier behind ChickDetect: how I trained it and how it performs on photos it never saw.

96.6%Test accuracy, all 5 classes
95.5%On the 4 dropping classes
  1. 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.

  2. 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).

  3. 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.

  4. Test

    Scored as the exported file, the same one the app runs, on 1,572 held-out test photos.