Research On Image Panoptic Segmentation Based On Neural Network Fusion

As a new task in the field of computer vision image segmentation,image panoramic segmentation task combines the detailed division of image foreground object feature information in the instance segmentation task and the in-depth understanding of image pixel level background semantic information in the semantic segmentation task.This paper studies the existing panoramic segmentation network algorithm based on semantic segmentation and instance segmentation,summarizes and explores a more concise network fusion method,which greatly reduces the amount of network parameter calculation and improves the panoramic segmentation speed of the model on the basis of ensuring the accuracy of image segmentation.Based on the basic panoramic segmentation algorithm of dual network fusion,in order to optimize the sub structure of each region of the network,reduce the overall parameters through the shared structure and ensure the high-level panoramic segmentation quality of the image,this paper proposes a method of semantic and instance segmentation network fusion assisted by group convolution,The following research work has been carried out:1.Based on the panoptic deeplab network model,this paper makes an in-depth analysis,and transforms the two branch decoding network of semantic segmentation and instance segmentation.By means of group convolution,the semantic and instance feature information obtained from the double ASPP spatial pyramid pooling structure is integrated into a unified single branch decoding network with non shared parameters,so as to reduce the amount of decoding network parameters and improve the network speed.At the same time,different multi-scale information obtained in the coding backbone network is fused with the top and bottom of the block convolution,so as to achieve the effect of integrating the global information of the image in different fields and optimizing the detail features.In the case of ensuring the optimal expression of semantic and instance information,the model becomes more rapid and concise in the process of network integration.2.By refining and classifying the hole convolution of different receptive fields,the semantic background features and instance foreground features of different regions are extracted,and the pyramid pooling structure of hole convolution space with different expansion rates is fused.The ASPP structure of semantic branch records more extensive background information,and the ASPP structure of instance branch records more obvious foreground information with edge features.And through comparative experiments to explore the influence of different expansion rate of double branch hole convolution on its feature expression,determine the optimal model structure,in order to achieve more precise image panoramic segmentation effect.Comparing the optimization model with other mainstream panoramic segmentation methods,the experimental results show that this paper can achieve the task of panoramic segmentation more concisely,and at the same time,it also ensures the very excellent panoramic segmentation effect.