A native Android Japanese dictionary app built using Kotlin and Jetpack Compose. It also contains a TFLite handwriting recognition model built using Python 3 and TensorFlow.
The ml folder contains the code for generating and training the dataset, the descriptions of each file can be seen as follows. This project can be run by importing the entire directory into android studio.
batch.sbatch
This runs the run.py file on the university server which uses powerful GPUs to speed up calculation.
convertotflite.py
This converts from a .keras file format into a .tflite format which allows it to be run on a mobile device.
createlabels.py
This creates a kotlin file that maps integers 0 to 6511 to their allocated Kanji.
datasetgenerator.ipynb
This was used for testing different image editing algorithms as well as generating the loss and accuracy graphs.
main.py
This was used to copy files from the original dataset into a usable format.
remove_unwanted.py
This is used to remove characters that we do not want to recognise, such as alphanumeric characters.
resize_all.py
This is used to generate the entire dataset including modifying images and normalising the amount of data that is in each class.
run.py
This contains the code to create and train the neural network.
testing.ipynb and testmodel.ipynb
This contains the code to test the model after training is completed, the two files exist to test the models on different devices.
The model above is then placed into a custom built android app, the results of which can be seen below.
