| 摘 要: 针对遥感图像目标检测因目标尺寸小和类别不均衡而导致的误检、漏检和置信度低的问题,提出了一种基于深度学习的目标检测方法。该方法的核心在于提出细粒度上下文模块构建主干网络,设计特征分化结构以增强对应尺寸目标的特征表达,定义自适应双重焦点损失函数替换交叉熵损失函数。提出的方法命名为细粒度特征分化网络。通过在RSOD、DOTA-1.0和VisDrone-2019三个公开的遥感图像数据集上进行实验,发现细粒度特征分化网络的mAP@0.5比基准模型YOLOv5s的相应指标分别提高了3.9百分点、4.2百分点和4.8百分点。实验结果表明,所提方法能够进行高效的遥感图像目标检测。 |
| 关键词: 深度学习;遥感图像;小目标检测;类别不均衡;焦点损失 |
|
中图分类号: TP391
文献标识码: A
|
|
| Target Detection in Remote Sensing Images Based on Deep Learning |
|
WU Jinda, LI Qiang
|
(School of Computer, Hangzhou Dianzi University, Hangzhou 310018, China)
1158587163@qq.com; hzlee@hdu.edu.cn
|
| Abstract: In response to the issues of false detection, missed detection, and low confidence in target detection of remote sensing images caused by small target sizes and class imbalance, this paper proposes a deep learning-based target detection method. The core of this method lies in the introduction of a Fine-Grained Context module to construct the backbone network, designing a feature differentiation structure to enhance the feature representation for targets of corresponding sizes, and defining an adaptive dual-focal loss function to replace the cross-entropy loss function. The proposed method is named Fine-Grained Feature Differentiation Network. Experiments conducted on three public remote sensing image datasets: RSOD, DOTA-1. 0, and VisDrone-2019, show that the mAP @ 0. 5 of the Fine-Grained Feature Differentiation Network is higher than the baseline model YOLOv5s by 3.9 percentage points, 4.2 percentage points, and 4.8 percentage points, respectively. The experimental results demonstrate that the proposed method is capable of efficient target detection in remote sensing images. |
| Keywords: deep learning; remote sensing images; small target detection; class imbalance; focal loss |
