U-Net with Dense Encoder, Residual Decoder and Depth-wise Skip Connections
Weiqin Ying, Junhui Li, Yu Wu, Kaijie Zheng, Yali Deng, Jiachen Li · 2020
For tasks like medical image segmentation and understanding, U-Net is one of the most prominent convolutional neural networks (CNNs) in recent years. Most of the models for image segmentation today are the variants of the classical U-Net. By applying some improvements on the convolution blocks and skip connections in U-Net, this paper proposes a dense-residual depth-wise U-Net (DR-DW U-Net). The DR-DW U-Net aims at extracting more useful features and alleviating the pressure during gradient descent. It adopts dense blocks in the left encoding path while using residual blocks in the right decoding path. In addition, the skip connections of DR-DW U-Net are injected with additional convolution blocks in the form of depth-wise convolution. The experimental results on two medical segmentation datasets indicate that the DR-DW U-Net achieves competitive performances and notable improvements over the classical U-Net as well as some well-designed variants of U-Net.