单阶段实例分割——从局部到整体的网络结构研究综述

Single-stage instance segmentation:a review of network structure research from local to global

单阶段实例分割是近年来深度学习领域的研究热点,其通过将目标检测和目标分割并行的方式实现图像的实例级分割,该方法目前已被广泛应用于图像目标分割领域。首先,阐述了单阶段实例分割基本原理。然后,从局部和整体2个方面对单阶段实例分割的网络结构进行梳理,在局部网络结构方面,从特征提取、特征融合、特征预测3个方面进行归纳,其中,在特征预测部分,按照有锚框到无锚框的思路对目标边界框的生成方式进行分类,按照全局掩膜到局部掩膜的思路对目标掩膜的表示方式进行分类,全局掩膜包括原型系数方法、目标位置方法和目标边界方法,局部掩膜包括目标轮廓方法、目标位置方法和目标特征方法;在整体网络结构方面,对22个主流的网络结构进行总结。接着,归纳了单阶段实例分割在医学图像分割、视频图像分割、遥感图像分割等应用领域的发展现状。最后,对单阶段实例分割的发展方向进行展望。

Single-stage instance segmentation is a hot research topic in the field of deep learning in recent years,in which the instance-level segmentation of images is realized by paralleling methods of object detection and object segmentation.This method has been widely used in the field of image object segmentation.Firstly,the basic principle of single-stage instance segmentation is described.Secondly,the network structure of single-stage instance segmentation is sorted from local and overall aspects.In terms of local network structure,the summarization includes three aspects:feature extraction,feature fusion,and feature prediction.Specifi-cally for the feature prediction,the generation method of the object boundary frame is classified according to the idea of anchor frame to non-anchor frame.The representation of object mask is classified according to the idea of global mask to local mask.The global mask methods include prototype coefficient method,object position method,and object boundary method,while the local mask methods include object contour method,object position method,and object feature method.In terms of the overall network struc-ture,the 22 mainstream network structures are summarized.Then,the development status of single-stage instance segmentation in medical image segmentation,video image segmentation,remote sensing image segmentation,and other application fields are sum-marized.Finally,the development directions of single-stage instance segmentation are prospected.