Image Panoptic Segmentation Based On Convolutional Neural Networks

In recent years,deep learning has been widely used in var:ious fields of computer vision,such as image detection and image semantic segmentation.A large number of practices and research have shown that deep convolutional neural networks can effec-ti'vely improve the precision of various fields of computer vision.The image panoptic segmentation task is to give each pixels in the image a unique semantic category and instance identification number.It is an important frontier problem of scene understanding.It can be widely used in robot visual perception,multitarget detection and automatic driving.Based on the convolutional neural network structure,this paper builds a network framework for panoptic segmentation,and discusses related issues.Based on the feature pyramid network structure,this paper build a new end-to-end panoptic segmentation network,which can obtain the prediction results in a single model.In order to explore the occlusion problem between different objects in the panoptic segmentation,this paper proposes a spatial hierarchical ranking module based on artificial prior and convolutional layer respectively,gets a new spatial priority ranking score that different from object detection score,and reorder according to this score.Thus it can solve the occlusion problem between different instance objects.This method can greatly im-prove prediction accuracy under different base models.Then,this paper analyze the problem of edge roughness in the panoptic segmentation.For the defects of the existing evaluation metric,the edge-friendly segmentation metrics are designed to reflect the performance of edge segmentation.This paper do experiments under the MS COCO 2018 panoptic segmentation training and validation dataset.The panoptic quality and others are set as the evaluation met-ric.Results show that the network structure designed can achieve excellent predictions and the spatial hierarchical ranking module can improve PQ largely under differen-t base models.By using the spatial hierarchical ranking methods and other methods,our methods win the MS COCO 2018 panoptic segmentation challenge,and get better results than others.Besides,the experiment results verify the effectiveness of the new metrics and provide the foundation for the follow-up research.