| 地面小目标快速检测算法研究 |
| |
| 引用本文： | 欧阳乐诚,王华力. 地面小目标快速检测算法研究[J]. 信号处理, 2019, 35(12): 1952-1958. DOI: 10.16798/j.issn.1003-0530.2019.12.003 |
| |
| 作者姓名： | 欧阳乐诚 王华力 |
| |
| 作者单位： | 陆军工程大学通信工程学院 |
| |
| 基金项目： | 国家自然科学基金（61271354） |
| |
| 摘 要： | 由于地面小目标像素低、携带信息量少等特点,常规数据集上的目标检测算法直接用于地面小目标的检测,会出现漏检和错检现象。针对此问题,该文基于深度学习算法,提出将一种改进的YOLOv3算法应用于地面小目标检测。利用K-means算法对数据集中目标进行聚类分析,选取合适的anchor boxes数量和大小。通过增加一个检测小目标的特征图,建立特征融合目标检测层,改进YOLOv3网络检测尺度。在遥感数据集DOTA检测实验中,用改进的YOLOv3算法与YOLOv3算法进行对比实验,结果表明改进后的算法能有效检测地面小目标,查准率提高17.04%,查全率提高了6.44%;与Faster R-CNN算法对比,改进的YOLOv3算法的mAP提高了12.7%。在改进的YOLOv3算法训练时，加入dropout优化机制，其在测试集上的mAP得分又提高了3.24%。
|
| 关 键 词： | 深度学习 YOLOv3算法 小目标检测 |
| 收稿时间： | 2019-08-22 |
Research on fast Detection Algorithm for Ground Small Targets |
| |
| Affiliation: | College of Communication Engineering, Army Engineering University of PLA |
| |
| Abstract: | Due to the characteristics of low pixels and little information carried by small ground targets, the target detection algorithm on the conventional data set is directly applied to the detection of small ground targets, which will result in missing and wrong detection. Aimed at this problem, this paper proposed an improved YOLOv3 algorithm for ground small target detection based on deep learning algorithm. In order to select the appropriate number and size of anchor boxes, K-means clustering algorithm was used to analyze the targets in the dataset. By adding a feature map for detecting small targets, the feature fusion target detection layer was established, which improved the YOLOv3 network detection scale. On the detection experiment of remote sensing data set of DOTA, the improved YOLOv3 algorithm and the YOLOv3 algorithm were compared. The results show that the improved algorithm can effectively detect small targets on the ground, the precision is improved by 17.04%, and the recall rate is increased by 6.44%. Comparing with the Faster R-CNN algorithm, the mAP of the improved YOLOv3 algorithm is improved by 12.7%. When the improved YOLOv3 algorithm was trained and dropout optimization mechanism was added, its mAP score on the test set was improved by 3.24%. |
| |
| Keywords: | |
|
| 点击此处可从《信号处理》浏览原始摘要信息 |
|
点击此处可从《信号处理》下载免费的PDF全文 |
|
