Panoptic Segmentation Algorithm Based On Feature Information Interaction

Panoptic segmentation is one of the important research directions in the interdisciplinary field of computer vision and applied mathematics.At this stage,most Panoptic segmentation algorithms use a combination of semantic segmentation and instance segmentation.The two subtasks share basic network features.Due to the connection between the two subtasks,in order to strengthen the information exchange between the subtask branches and promote the semantic branch to accurately predict the stuff in the image,this paper mainly studies the panoptic segmentation algorithm of feature information interaction.The content is as follows:Firstly,based on the EncNet algorithm,a semantic segmentation algorithm that can solve the problem of imprecise segmentation of large objects and extract multi-scale contextual semantic information is proposed.The algorithm uses a dense expansion convolution pyramid structure,which is formed by cascading multiple expansion convolutions.While increasing the receptive field of the feature map,it can extract semantic information at multiple scales.In order to solve the problem that the model error rate increases due to BN normalization under small batch size,the algorithm uses GN normalization instead of the normalization method of the original algorithm.Experiments show the effectiveness of the proposed method.The mIoU in the dataset PASCAL VOC2012 reaches 84.9%.Secondly,this paper proposes a panoptic segmentation algorithm for feature information interaction.This algorithm contains two subtask branches:semantic segmentation and instance segmentation.The two subtask branches share the basic network,and the semantic segmentation branch uses an improved semantic segmentation algorithm.The instance segmentation uses the Mask RCNN network structure.In order to promote mutual learning between subtasks,this paper proposes a regional attention module,which uses the RPN network in instance segmentation to promote the prediction of the stuff segment in the image by the semantic segmentation branch.In the process,it also promotes better instance segmentation subnetwork Learning.Finally,the output fusion of semantic segmentation and instance segmentation is used to obtain the result of panoptic segmentation.In order to make the network learn better,this paper uses a joint training method,where the loss includes two parts:semantic segmentation loss,instance segmentation loss,and panoptic segmentation loss as a weighted sum of the two part losses.The experimental results show that the panoramic segmentation algorithm of feature information interaction in this paper has high accuracy in the panoramic segmentation dataset.In the dataset COCO 2018 Panoptic,the PQ reached 39.8%,the RQ was 48.8%,and the SQ was 76.2%.In the dataset Cityscapes,the PQ reached 58.5%,the RQ was 71.9%,and the SQ was 77.8%.