Abstract:
Aiming at the overlapping phenomenon of background instances in panoptic segmentation, an attention-guided panoptic segmentation network (APSNet) based on the attention mechanism is proposed. APSNet is improved based on the UPSNet network. Firstly, a triplet attention mechanism is added between residual network and feature pyramid network, which can model spatial attention and channel attention at the same time. By learning channel and spatial feature information, the feature extraction ability of feature pyramid network is enhanced. Secondly, in the semantic segmentation part, a kind of semantic is added to enhance attention. By aggregating multi-level features and learning spatial feature information, the segmentation effect of semantic segmentation for background and foreground is improved. The comparison experiments show that the panoptic quality of APSNet is improved by 0.8%, and the instance level panoptic quality is improved by 2.7%. At the same time, it can reduce the probability of overlapping of background instances in panoptic fusion.