!! update ported some of the network to pytorch for it's easier to convert to ncnn. Check it here: mobile phone portrait matting

Hair_Segmentation_Keras

Implement some light weight hair segmentation network with keras which can be used on mobile devices easily.

Dataset

1. CelebA Face/Hair segmentation database
2. Figaro
3. LFW Part Labels

Model

1. [DeeplabV3plus]: MobileNetV2 as the encoder
2. PrismaNet: network architecture as described in the Prisma-AI blog
3. FastDeepMatting
4. PrismaNet + FastDeepMatting: base PrismaNet architecture plus the feathering block in fast deep matting

Training

Results

DeeplabV3plus

PrismaNet

FastDeepMatting

PrismaNet + FastDeepMatting

Matting methods used channel split operation which is unportable to CoreML as I wrote.

Serving

I have also use this model to predict hair color with tensorflow serving. Follow instructions bellow.

1. Use this scripts to python serving/keras_to_serving.py generate model used for tensorflow serving deployment.
2. Prepare tensorflow serving environments. Please refer to README.md

Todo

• port to iOS using CoreML Hair_Segmentation_iOS
• tutorial about how to serve the model using tensorflow serving and use the results for hair color prediction. Hair Color Predict Using TensorFlow Serving
• update to TensorFlow2.0 Keras API