A Novel Face Attribute Segmentation Algorithm

We propose a novel segmentation algorithm for face attribute segmentation with a fully convolutional encoder-decoder network. Our network is trained end-to-end on Helen dataset, and optimizes by minimizing two loss functions: the L1 loss and the negative-log-likelihood. Strategies such as transfer learning, dilated convolution, skip layer are used to improve network accuracy. We also add maximum connected region extraction (MCRE) to the output of the network, and show that these strategies have significantly improved the network performance. The experimental results show that the network has obtained F-Measure 0.872 on the Helen dataset, which yielding higher score in facial segmentation than previous methods. Moreover, our algorithm also achieves a good visual segmentation effect on other images outside the dataset, which demonstrate strong generalization performance of the proposed algorithm.