Top-Down Sampling Convolution Network for Face Segmentation

The paper adopts two different convolution sampling paths: from large scale to small scale sampling (top-down) and small scale to large scale sampling (bottom-up), and propose the top-down sampling convolution neural network for face segmentation (TDNN). On the LFW and the Helen dataset, it is demonstrated about the advantage of face segmentation by TDNN. In addition, the shared weight is added to each convolution integral, we propose TDNN with shared weight (TDNNSW). On the Helen dataset, TDNNSW with shared weight further improves the accuracy of face segmentation. Since TDNN is trained end-to-end, our model has advantageous properties such as less parameters and more rapid calculation for face segmentation.