Region-and-Attention Network for Semantic Segmentation

In this article, we address the task of semantic segmentation by aggregating the information of feature regions at different scales and modeling the dependency between regions by combining the attention mechanism. Instead of using global pooling to bundle information from each channel into a single value in the past, we propose a Region-and-Attention Network (RANet) to pay attention to local features and their global dependence at the regional level. Specifically, we use different pooling rates for the high-level output to obtain multi-scale information aggregates, and regard the multi-channel sequence contained in each pixel position in the pooling result as the global information integration of the corresponding region before the feature layer pooling. Unlike SENet, we perform a separate channel attention transformation for the regional information represented by each pixel position, and use one-dimensional convolution instead of multi-layer perceptrons. The final segmentation result is enhanced region-level prediction by fusing RANet outputs at different pooling scales. The proposed approach achieves advanced segmentation performance on two challenging common segmentation datasets, Cityscapes and Pascal VOC.