基于深度学习的实例分割研究进展

Survey of Instance Segmentation Based on Deep Learning

目标检测确定检测图像中目标对象所在区域及其类别,语义分割对检测图像实现像素级分类,实例分割可以定义为同时解决目标检测与语义分割问题,在分类的同时确定每个目标实例语义。实例分割网络在无人机驾驶、机器人抓取、工业筛检等领域具有重要应用意义,针对目前基于深度学习实例分割综述性文章的空白,对实例分割进展进行概述,按照单阶段实例分割与双阶段实例分割的分类对不同网络模型进行论述,重点介绍近两年网络框架的发展,总结各网络特点的同时提出未来发展方向。

Target detection determines the area and category of target objects in the detected image.Semantic segmentation implements pixel-level classification of detected images.Instance segmentation can be defined as solving the problem of target discovery and semantic segmentation at the same time,and the semantics of each target instance are determined during classification.Instance segmentation networks have important application importance in areas such as drone driving,robot gripping,and industrial screening.Aiming for a blank in the current review article based on deep learning instance segmentation,this article provides an overview of single-stage instance segmentation and double instance segmentation.The stage instance segmentation classification describes different network models,focuses on the development of network frameworks over the last two years,summarizes the characteristics of each network,and suggests future development directions.