摘要
卫星影像特定目标识别与定位可为军事侦察领域提供重要战略信息,如何平衡基于深度学习技术遥感目标检测算法的检测精度、参数量大小与检测效率成为当前研究热点。提出算法在Tiny-Yolo-V2算法基础上优化,利用深度可分离卷积概念解藕卷积层,减少模型参数量,提升检测效率。同时,结合特征金字塔思想,增加预测尺度,提高算法检测精度。在DOTA数据集上进行对比实验,结果表明相对Tiny-Yolo-V2算法,提出方法的mAP值提高了0.084,模型参数量减小了49%,检测效率提高了48%,验证了优化策略的有效性。
Identification and positioning of specific targets in satellite imagery can provide important strategic information in the field of military reconnaissance. How to balance the detection accuracy, parameter size and detection efficiency of remote sensing target detection algorithms based on deep learning technology has become a hot research topic. The algorithm proposed in this paper was optimized on the basis of the Tiny-Yolo-V2 algorithm. A deep separable convolution concept was usded to deconvolve the convolution layer to achieve the purpose of reducing the amount of model parameters and improving the detection efficiency. At the same time, the algorithm combines the idea of feature pyramid to achieve the effect of increasing the prediction scale and improving the detection accuracy of the algorithm. The comparison experiments on the DOTA dataset show that compared with the Tiny-Yolo-V2 algorithm, the mAP value of the proposed method is increased by 0.084,the model parameter amount is reduced by 49%,and the detection efficiency is increased by 48%,which verifies the effectiveness of the optimization strategies.
作者
张曼
叶曦
李杰
沈霁
ZHANG Man;YE Xi;LI Jie;SHEN Ji(Shanghai Aerospace Electronics Technology Research Institute,Shanghai 201109,China)
出处
《计算机仿真》
北大核心
2022年第2期17-22,97,共7页
Computer Simulation
关键词
卫星影像
目标检测
深度可分离卷积
特征金字塔
Satellite image
Object detection
Deep separable convolution
Feature pyramid
