PPNet : pooling position attention network for semantic segmentation

Semantic segmentation with attention module has made great progress in many computer vision tasks. However, attention modules ignore some boundary information. To explore a more comprehensive map of context features, we propose a pooling position attention network (PPNet) for semantic segmentation. Based on the Encoder-Decoder structure, we import attention modules into the encoder to enhance the correlation between deep information. Pooling cross attention module (PCAM) aims to weight deep semantic information and expands the feature recognition area, and pooling position attention module (PPAM) calculates the weighted features to generate features with strong semantic information. Finally, the enhanced deep features and shallow features are fused by decoder to enhance the dependency between pixels and to achieve better semantic segmentation. Experiments show that of our proposed PPNet is superior to other state-of-the-art models in the performance of segmentation accuracy on datasets PACSCAL VOC 2012 and Cityscapes.