无定形区特征增强全景分割算法

针对UPSNet全景分割算法在目标分割过程中存在对无定形区目标分割精度不高的现象,提出一种无定形区特征增强的UPSNet全景分割算法APS。该算法引入空洞卷积,采用自下而上的方式构建特征融合结构,使无定形区特征得以增强,解决无定形目标特征不显著问题,提高语义分割效果,进一步提高全景分割精度。经仿真测试,该算法对道路、草地等无定形目标分割效果均有提高,在COCO数据集上的检测结果与UPSNet相比,SQ值提高4.75%,适合应用于无人车和移动机器人等场景。

Amorphous Region Feature Enhanced Panoptic Segmentation Algorithm

Aiming at the problem of ignoring the target of amorphous area in the target segmentation process of UPSNet panoptic segmentation algorithm, this paper proposes an amorphous region feature enhanced UPSNet (APS) panoptic segmentation algorithm. This algorithm introduces dilated convolution and uses a bottom-up approach to build a feature fusion structure, which enhances the features of the amorphous region, solves the problem of insignificance of the amorphous target features and improves the effect of semantic segmentation and the accuracy of panoptic segmentation. The algorithm improves the segmentation results of amorphous targets such as roads and grasslands. Test result on the COCO 2017 dataset shows that SQ value is increased by 4.75% compared to the original UPSNet panoptic segmentation algorithm, which is suitable for panoptic segmentation environments such as autonomous vehicles and mobile robots.