激光与光电子学进展, 2020, 57(12): 120005, 网络出版: 2020-06-01

深度学习目标检测方法及其主流框架综述

Review of Deep Learning Based Object Detection Methods and Their Mainstream Frameworks
作者单位

1贵州大学现代制造技术教育部重点实验室, 贵州 贵阳 550025

2贵州大学机械工程学院, 贵州 贵阳 550025

摘要
目标检测作为机器视觉中重要任务之一,是人工智能体系中一个具有重要研究价值的技术分支。对于卷积神经网络框架、anchor-based模型和anchor-free模型三个主流的目标检测模型进行梳理。首先,综述了主流卷积神经网络框架的网络结构、优缺点以及相关的改进方法;其次从one-stage和two-stage两个分支对anchor-based类模型进行深入分析,总结了不同目标检测方法的研究进展;从早期探索、关键点和密集预测三部分分析anchor-free类模型。最后对该领域的未来发展趋势进行了思考与展望。
Abstract
As one of the important tasks in machine vision, object detection is a technology branch with important research value in artificial intelligence systems. The three mainstream object detection models of convolutional neural network framework, anchor-based model, and anchor-free model are analyzed. First, the network structure and the advantages and disadvantages of the mainstream convolutional neural network framework, and the related improvement methods are reviewed. Second, the anchor-based model is deeply analyzed from one-stage and two-stage branches, and the research progresses of different object detection methods are summarized. The anchor-free model is analyzed from three parts: early exploration, key points, and intensive prediction. Finally, the future development trend of the field is considered and prospected.
00 11