A Web API for automatic background removal using Deep Learning. App is made using Flask and deployed on Heroku.
👉 https://portrait-me.herokuapp.com/
CNN Architecture - U-Net with Residual connections
Parameters - 8.9M
Trained on - 64,115 Images
validated on - 2693 Images
batch_size = 32
img_size = (256,256)
Trained for - 13 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 - 0.038
Validation loss - 0.056
Training accuracy - 0.935
Validation accuracy - 0.907
Training meanIOU - 0.817
Validation meanIOU - 0.787