%0 Journal Article
%T 基于深度学习的有向目标检测研究进展
Research Progress of Directed Object Detection Based on Deep Learning
%A 罗鸿亮
%A 王武鑫
%A 叶轩宇
%A 朱晟翔
%A 白永强
%J Journal of Image and Signal Processing
%P 258-270
%@ 2325-6745
%D 2024
%I Hans Publishing
%R 10.12677/jisp.2024.133022
%X 目标检测是计算机视觉领域的重要任务，旨在从图像或视频中准确地识别和定位出现的目标物体。但是，普通目标检测算法往往难以处理旋转、带有方向信息的物体。针对此问题，诞生了许多专门用于有向目标检测方法，这些方法提供了更好的面向目标的空间表达，在图像处理方面取得了重大进展。有向目标检测的基本原理是将检测目标的旋转方框和倾斜角以及特征先界定再表示。本文综述了有向目标检测现阶段的国内外研究现状，根据有无锚框将当前基于深度学习的有向目标检测方法分为了基于锚框的一阶段方法、基于锚框的二阶段方法和无锚框方法3类方法进行归纳分析，并从优缺点、骨干网络、适用场景和数据集等方面进行了对比。最后，对有向目标检测方法的发展前景和研究方向进行了展望。
Object detection is an important task in the field of computer vision, aimed at accurately identifying and locating objects appearing in images or videos. However, conventional object detection algorithms often struggle to handle objects with rotation or directional information. To address this issue, many specialized methods for oriented object detection have been developed, offering better spatial representation tailored to the objects and achieving significant progress in image processing. The basic principle of oriented object detection is to define and represent the rotation bounding box, tilt angle, and features of the detected objects. This paper reviews the current research status of oriented object detection both domestically and internationally, categorizing current deep learning-based oriented object detection methods into three types: one-stage methods based on anchor boxes, two-stage methods based on anchor boxes, and anchor-free methods, based on whether anchor boxes are used or not, and conducts a comparative analysis from aspects of advantages, disadvantages, backbone networks, applicable scenarios, and datasets. Finally, the paper discusses the prospects and research directions of oriented object detection methods.
%K 有向目标检测，计算机视觉，深度学习，综述，卷积神经网络
Oriented Object Detection
%K Computer Vision
%K Deep Learning
%K Overview
%K Convolutional Neural Network
%U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=91228