handwritten-bengali-charecter-recognition
hii all,
This is a repo on hand written bengali charecter recognition . I have used transfer learning on VGG16 model. Libraries used is maiinly keras and tensorflow.
I have achieved 93% accuracy on the validation dataset with a 21% loss. In future much accuracy can be achieved by further fine tuning the code. The accuracy achieved is one of the highest using the dataset.
future scope:
1: make a ui for validation of the model using images from outside. 2: accuracy can be further increased by fine tunung the optimizers as ell as the model.
steps to use: 1: download the dataset from https://www.dropbox.com/s/55bhfr3ycvsewsi/CMATERdb%203.1.2.rar 2: unzip it in the directory you want 3: modify the code for the directory 4: train the model.
MAIN STEP: install all the dependency for the code to work.
1: Keras 2: Tensorflow 3: PILLOW
System requirement:
I have trained the following model on
AMD Ryzen 3 2240g NVIDIA gtx 750 ti as gpu
you can install tensorflow gpu if you have a graphics card.
to run model.py run as following
python model.py --train for training python model.py --test for testing