PaFPN-SOLO: A SOLO-based Image Instance Segmentation Algorithm

In order to improve the image instance segmentation algorithm due to the long propagation path of the underlying location information and the slow speed of convolutional operations in the process of capturing long-distance dependencies due to low computational efficiency, a PaFPN-SOLO algorithm is proposed in this paper. By adding Non-local operation to the ResNet backbone, the feature information of the image in the feature extraction process is better preserved; by using the bottom-up path augmentation method, more accurate position information is extracted on the lower feature layers, which not only improves the feature structure localization ability of the network model, but also shortens the information propagation path between the feature layers. The experimental results show that the algorithm in this paper has good segmentation effect on both COCO2017 and Cityscapes datasets, and the average segmentation accuracy reaches 56% and 47.3%, respectively, which improves 4.4% and 7.4% compared with the original SOLO network, respectively.