Pyramid attention network for indoor scene segmentation

With the rapid development of artificial intelligence technology, indoor scene segmentation is more and more widely used in robot vision, security monitoring, firefighting, and other fields. Usually, indoor objects have problems such as easy mutual occlusion and dense layout, thus requiring high accuracy of network segmentation. However, while improving network segmentation accuracy, the network operation speed is often sacrificed, and the recognition effect is challenging to meet the requirements. Since the existing indoor scene network recognition ability is not strong and the segmentation accuracy is not high. This paper proposes a new Pyramid Attention Network (PANet), which improves the previous multi-scale feature extraction method, and designs the Scene Attention Module (SAM) so that the network can combine the powerful global parsing capability of Atrous Spatial Pyramid Pooling (ASPP) layer with the Scene Attention Module. PANet first extracts feature maps at different scales and then applies the spatial attention mechanism to the feature maps so that the network forms an enhanced representation of information in each region of the feature maps. PANet achieves 35.12% mIoU and 37.96fps segmentation speed on NYU V2, an indoor scene dataset produced by New York University, which can be effectively used for real-time image processing.