A novel deep instance feature chain learning network for panoptic segmentation(INFNet)was developed to solve the problem of failure of target boundary segmentation caused by insufficient instant fea. ture extraction in panoptic segmentation. This network consisted of a basic chain unit, whose functions were divided into two types, feature holding chain and feature enhancement chain, based on the different methods of processing feature information by the unit structure. The feature-holding chain represented the input stage of the extraction of a chain network feature, in which the integrity of the input information was guaranteed, and then this feature was transmitted to the feature-enhancement chain structure. The feature-enhancement chain increased the network depth and improved the feature extraction ability through its ex. tension. INFNet could obtain adequate edge feature information and improve segmentation accuracy, ow. ing to the robust depth-stacking characteristics. The experiment results for the MS COCO and Cityscapes datasets showed that our INFNet was superior to similar existing methods in terms of segmentation accura. cy. Compared to the Mask RCNN instance segmentation structure widely used in panoptic segmentation networks, the segmentation accuracy of INFNet increased by up to 0. 94%.