A Web API for automatic background removal using Deep Learning

Automatic_Background_Remover

A Web API for automatic background removal using Deep Learning. App is made using Flask and deployed on Heroku.

Automatic_Background_Remover

CNN Architecture - U-Net with Residual connections
Parameters - 2.2M
Trained on - 153,947 Images
validated on - 2693 Images
batch_size = 32
img_size = (256,256)
Trained for - 4 epochs 
Training time - 80min/epoch on GPUs by Google Colab.

Datasets used for training:

The model is trained using modified version of U-NET (https://arxiv.org/abs/1505.04597) Architecture first presented by Olaf Ronneberger, Philipp Fischer, Thomas Brox in 2015.
I have added Residual skip connections in U-NET Model which makes it more robust.

I can’t put model architecture here because of its huuge size. view here.

Training loss - .112
Validation loss - .134

 

 

 

To finish reading, please visit source site