Fig. 2

The architecture for the deep neural networks in our proposed system. The deep neural networks are composed by the deep autoencoder for high-level feature extraction and the softmax classifier for soft mask generation and was trained by the greedy layer-wise training method