基于深度学习的实例分割研究综述

深度学习在计算机视觉领域已经取得很大发展，虽然基于深度学习的实例分割研究近年来才成为研究热点，但其技术可广泛应用在自动驾驶，辅助医疗和遥感影像等领域。实例分割作为计算机视觉的基础问题之一，不仅需要对不同类别目标进行像素级别分割，还要对不同目标进行区分。此外，目标形状的灵活性，不同目标间的遮挡和繁琐的数据标注问题都使实例分割任务面临极大的挑战。本文对实例分割中一些具有价值的研究成果按照两阶段和单阶段两部分进行了系统性的总结，分析了不同算法的优缺点并对比了模型在COCO数据集上的测试性能，归纳了实例分割在特殊条件下的应用，简要介绍了常用数据集和评价指标。最后，对实例分割未来可能的发展方向及其面临的挑战进行了展望。

A survey of instance segmentation research based on deep learning

Deep learning has made great progress in the field of computer vision. Although instance segmentation research based on deep learning has only become a research hotpot in recent years, relevant techniques can be widely used in the fields of autonomous driving, complementary medicine and remote sensing imaging. Instance segmentation, as one of the fundamental problems of computer vision, requires not only pixel-level segmentation of different classes of targets, but also differentiation of different targets. In addition, the flexibility of target shapes, the occlusion between different targets and the tedious data annotation problems all make the instance segmentation task extremely challenging. In this paper, firstly, some valuable research results in instance segmentation are systematically reviewed according to two-stage instance segmentation and one-stage instance segmentation. Secondly, the advantages and disadvantages of different algorithms are analyzed and the testing performance of different models on the COCO dataset is compared. In addition, the applications of instance segmentation under special conditions are summarized, and common datasets and evaluation metrics are briefly introduced. Finally, the possible future directions of instance segmentation and the challenges it faces are prospected.