Skip to content
Dinosaur species image classification with TensorFlow Lite and Flutter.
Dart Ruby Swift Kotlin Objective-C
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
android
assets
ios
lib
.gitignore
.metadata
README.md
demo.mp4
pubspec.lock
pubspec.yaml
screenshot.png

README.md

DinoFinder

A machine learning model for native mobile devices to perform dinosaur image classification using TensorFlow Lite and Flutter.

Supported species:

  • Ankylosaurus
  • Brachiosaurus
  • Dilophosaurus
  • Dimetrodon
  • Iguanodon
  • Protoceratops
  • Pteranodon
  • Spinosaurus
  • Stegosaurus
  • Triceratops
  • Tyrannosaurus
  • Velociraptor

Screenshot

View Demo

Install dependencies:

flutter pub get

Run:

flutter run

Methods:

The camera from a user's mobile device is used to capture a square, center-cropped image. This image is rescaled with bicubic interpolation down to a 224x224px image. The pretrained model performs a prediction on this data as a standarized 3-channel RGB matrix. The prediction result with the highest confidence score is then overlaid onto the user interface.

Image scraper:

To collect training data-set for dinosaur image classification, the image scraper that efficiently collects image data is implemented. The image scraper link is as follows: Google Image Scraper

Machine Learning(ML) model

The ML model uses mobilenet provided by Tensorflow. The following link has the ML model. Dinosaur classification model

Limitations:

We aim to perform dinosaur classification across a range of artist renderings for a small set of well-known species of dinosaurs. Acquiring the training data, filtering out the best-representational imagery, and generalizing the features between results is not a simple task. Our current model was trained on just a few hundred samples for each label type, limiting the accuracy of the model. We hope to expand our dataset over the coming months to improve the model accuracy using a wider collection of publicly available renderings on the web, then publish the application as a free download for iOS and Android devices.

Credits:

This work was completed by Youngjun Choi and Ryan Bell.

You can’t perform that action at this time.