| 解博,朱斌,樊祥,张宏伟,马旗,张扬.基于SSD卷积网络的航拍图像目标检测方法[J].光电子激光,2019,30(4):407~414 |
| 基于SSD卷积网络的航拍图像目标检测方法 |
| Object detection method based on SSD convolution network in aerial images |
| 投稿时间：2018-09-06 |
| DOI： |
| 中文关键词: 目标检测 深度学习 航拍图像 梯度聚类 SSD(single shot detector) |
| 英文关键词:object detection deep learning aerial imagery gradient clustering single sho t detector (SSD) |
| 基金项目:国家自然科学基金(61271376)资助项目 (国防科技大学 脉冲功率激光技术国家重点实验室,安徽 合肥 230037) |
|
| 摘要点击次数: 725 |
| 全文下载次数: 0 |
| 中文摘要: |
| 基于深度学习的目标检测算法是目前目标检测领域 最流行的算法,但是由于硬件条件的限制,算法输入图像的尺寸受到限制。对于大尺寸的航 拍图像,通常先采用滑窗法提取区域,再对提取的 区域进行检测,极大地降低了算法的检测速度。针对这一问题,本文根据航拍图像中人造物 体含有 大量边缘的特点,提出了一种基于深度学习的梯度聚类目标检测算 法,并阐述了其模型结构与工作原理,然后通过151张航拍图像数据 集测试,对比评估了梯度聚类 SSD方法与滑窗SSD方法在航拍图像检测上的检测精度和检测速度。结果表明:梯度聚类SSD 方 法的FPS(Frames Per Second)为0.499,SPF(Seconds Per Frame )为2.00,mAP(mean Average Precision) 为46.93,相比滑窗SSD方法,在损失11.72%的 检测精度的条件下,FPS提高了64.69%(SPF提高了40.40%),验证了所提出算法的有效性。 |
| 英文摘要: |
| Object detection method based on deep learning is the most popular object detection algorithm in the field of object d etection currently.Due to limitations of hardware conditions,the size of the input image for the algorithm is limited.Aerial imageries usually h ave a large size.So the sliding windows method is usually used to extract the areas first,and then the extracted ares are detected,which greatly influence the detection speed of the algorithm.This p aper provides a object detection method that is the single shot detector (SSD) based on edges gradient clustering of objects in aerial images to solve the problem.The model architecture and its working principle we re deeply expounded in this paper.Then,through the test in a dataset of 151aerial images, the detection performance of gradient clustering SSD and sliding windows SSD were evaluated an d compared in object detection speed and precision.Experimental results show tha t the gradient clustering SSD is 0.499in frames per second (FPS),2.00in senconds per frame (SPF) and 46.93in m ean average precision (mAP).Compared with the sliding windows SSD,the FPS has respec tively improved by 64.69% (the SPF has respectively raised by 40.40%),but the mA P has respectively declined by 11.72%.Thus the effectiveness of the proposed method is verified. |
| 查看全文 下载PDF阅读器 |
| 关闭 |
|
|
|
|
|