Doodlifier is a program to classify doodles and alphabet drawn by the user into different classes with their corresponding confidence percentages.
- Install tensorflow
- Install requirements
- Run
main.py
Uses existing model and class names by default.
To customize doodles, delete files under model and update doodle.txt with reference to available.txt, and run main.py.
Doodlifier uses a Convolutional Neural Network to extract features, enhance them and then using softmax returns probabilities for each class label. An accuracy of roughly 95 percentage was attained.
-
From doodle.txt, the required class npy files are downloaded from Google's Quickdraw datasets.
Check available.txt for a list of available classes.
-
They are randomized to try eliminating intialization bias and classes are one hot encoded and then trained and tested.
-
Images are resized to 28x28 and preprocessed.
-
These images are passed to the model and then class probabilities are returned.
-
The model is then saved along with class names.
Doodles
- https://github.com/googlecreativelab/quickdraw-dataset
- https://quickdraw.withgoogle.com/data/
- https://storage.googleapis.com/quickdraw_dataset/
- https://en.wikipedia.org/wiki/List_of_datasets_for_machine-learning_research#Handwriting_and_character_recognition
- https://faroit.com/keras-docs/1.2.0/
- https://github.com/tensorflow/docs
- IEEE Papers
- https://ieeexplore.ieee.org/document/9734453
- https://ieeexplore.ieee.org/document/9137933
- https://ieeexplore.ieee.org/document/9537097
- https://ieeexplore.ieee.org/document/7903730
- https://ieeexplore.ieee.org/document/8079572
- https://ieeexplore.ieee.org/document/8203926
- https://ieeexplore.ieee.org/document/9213619
- https://ieeexplore.ieee.org/document/8365201
- https://ieeexplore.ieee.org/document/9675509
- Muliple detections with the help of YOLO
- Better datasets for English Characters and digits
- Add more models