Dense Convolutional Networks for Semantic Segmentation

Recent studies have greatly promoted the development of semantic segmentation. Most state-of-the-art methods adopt fully convolutional networks (FCNs) to accomplish this task, in which the fully connected layer is replaced with the convolution layer for dense prediction. However, standard convolution has limited ability in maintaining continuity between predicted labels as well as forcing local smooth. In this paper, we propose the dense convolution unit (DCU), which is more suitable for pixel-wise classification. The DCU adopts dense prediction instead of the center-prediction manner used in current convolution layers. The semantic label for every pixel is inferred from those overlapped center/off-center predictions from the perspective of probability. It helps to aggregate contexts and embeds connections between predictions, thus successfully generating accurate segmentation maps. DCU serves as the classification layer and is a better option than standard convolution in FCNs. This technique is applicable and beneficial to FCN-based state-of-the-art methods and works well in generating segmentation results. Ablation experiments on benchmark datasets validate the effectiveness and generalization ability of the proposed approach in semantic segmentation tasks.