摘要
近年来,随着科学技术的快速发展和硬件设备的不断迭代,人工智能在各种领域(如安防监控、医疗辅助、健康诊断、智能推荐、遥感监测、目标定位等)都得到了广泛的应用.随着人们对智能处理任务的要求逐步提高,人工智能算法所需要理解的语义信息层次和输出数据精准度要求也步步攀升.因此,像素级语义理解任务也因其精准度要求远高于图像级理解而越来越受到重视.与图像级理解相比,像素级语义理解具有输出数据量大、逐像素输出精度高的优点,相应地其难度也更大,内部成因更值得关注与研究.为此本文从信息度量的角度出发,结合像素级语义理解任务的特有属性,给出了像素级语义理解任务的定义与优化目标,进一步依据实际任务的特性从初始定义衍生出像素级语义分类和像素级语义回归两类任务;随后分别讨论了在这两类任务中优化目标的退化和演变,并通过详尽的调研总结了常见像素级语义理解任务的发展现状;紧接着探究了当前像素级语义理解的难点和未来发展方向,针对亟待解决的问题给出了深入的分析思考以及可行的解决方案;最后重点反思了后深度学习时代像素级语义理解乃至人工智能领域所面对的机遇与挑战,提出知识的方向指导和数据的优化驱动是未来人工智能发展的重点关注目标.本文意图从像素级语义理解的定义与发展现状出发,延伸出对当前工作的思考以及对整个领域的反思,强调整个领域面临的风险;在介绍像素级语义理解基础认知的同时对相关技术的发展方向和路径进行深入的思考与深度的展望.
With the rapid developments in science and technology and the continuous iteration of hardware equipment, artificial intelligence is being widely used in various fields(such as security monitoring, medical assistance, health diagnosis, intelligent recommendation, remote sensing monitoring, and target location). With the gradual expansion of user requirements for intelligent processing tasks, semantic understanding levels and accuracy requirements of intelligent algorithms are also gradually increasing. As such, the pixel-level task of semantic understanding, which requires much higher accuracy than image-level understanding, is attracting increasing attention. Compared to image-level understanding, in addition to the large amount and high precision of pixel-by-pixel output, the internal challenges in pixel-level semantic understanding merit greater attention and research. From the information measurement perspective, and given the unique attributes of pixel-level semantic understanding, in this paper, we present the definition and optimization goal of pixel-level semantic understanding, and derive pixel-level semantic classification and pixel-level semantic regression from the original definition based on the characteristics of actual tasks. We then discuss the degradation and evolution of optimization objectives in these two subcategories, respectively. Based on a detailed investigation, we summarize the development and current status of typical tasks in pixel-level semantic understanding. We consider the difficulties and future development direction of the current pixel-level semantic understanding, perform an in-depth analysis, and suggest feasible solutions to problems that require urgent solutions. Finally, we focus on the opportunities and challenges faced by pixel-level semantic understanding, and even artificial intelligence, in the post-deep-learning era and argue that knowledge guidance and data-driven optimization are fundamental objectives for the future development of artificial intelligence. From th
作者
李学龙
赵致远
Xuelong LI;Zhiyuan ZHAO(School of Computer Science and School of Artificial Intelligence,Optics and Electronics(iOPEN),Northwestern Polytechnical University,Xi'an 710072,China)
出处
《中国科学：信息科学》
CSCD
北大核心
2021年第4期521-564,共44页
Scientia Sinica(Informationis)
基金
科技部重点研发计划(批准号:2018YFB1107400)
国家自然科学基金(批准号:61871470)资助项目。
关键词
像素级语义理解
人工智能
深度学习
分类
回归
pixel level semantic understanding
artificial intelligence
deep learning
classification
regression
