Fully convolutional network with attention modules for semantic segmentation

Fully convolutional network is a powerful end-to-end model for semantic segmentation. However, it performs prediction pixel by pixel to pose weak consistency on intra-category. This paper proposes fully convolutional network with attention modules for semantic segmentation. Based on the framework of fully convolutional network, the post-processing attention module and skip-layer attention module are introduced to enhance the relevancy among pixels. Post-processing attention module is to calculate the similarity among pixels to obtain global information. Skip-layer attention module is designed to combine semantic information from a deep, coarse layer with contour information from a shallow, fine layer to produce the feature with high resolution and strong semantic information. Loss function, obtained by cross-entropy between estimated probability and label, is to optimize the network. Extensive experiments demonstrate that the proposed approach is superior to DeepLab and other models in performance of mean IoU with moderate computational complexity

Automated summary

The paper proposes a fully convolutional network with attention modules for semantic segmentation, which enhances pixel relevancy through post-processing and skip-layer attention, optimizes the loss function, and improves performance compared to other models.