A Deep Learning Approach to Hair Segmentation and Color Extraction from Facial Images

model

Segmentation : U-Net(with depthwise conv) + VGG structure => Background/Hair/Face ( Not for Mobile ) -> + bald segmentation + hair color segmentation

Dataset prepare

cluster with slic -> make k-means (20) and extract hair color ( use celeb A )

=> Classify hair color with random forest ( feeding histogram )

Labeled Faces in the Wild (LFW) dataset + Figaro 1K + celebA (use for hair color )

Used loss

Iou + pixel accuracy

Output