摘要
针对低算力硬件环境下,常规遥感影像目标检测方法难以正常运行的问题,提出一种轻型遥感影像检测模型。以深度可分离卷积核与平均池化下采样并联结构搭建特征提取结构,然后在特征提取结构末端连接双层轻型特征强化金字塔,最后在检测输出端口使用改进非极大值抑制算法输出最佳检测框。以多个开源遥感影像数据集构建测试数据集,同时使用改进的k-means++算法进行锚点框聚类。实验结果表明,本文模型在平均精度均值方面显著优于对比模型,同时能够在低算力硬件中开展实时级检测,具有实际应用价值。
Aiming at the problem that conventional remote sensing image target detection methods are difficult to operate normally in a low-computing power hardware environment,a lightweight remote sensing image detection model is proposed.The feature extraction structure is built with a parallel structure of depth-separable convolution kernel and average pooling downsampling,and then a double-layer lightweight feature enhancement pyramid is connected at the end of the feature extraction structure.Finally,an improved non-maximum suppression algorithm is used at the detection output port to output the best detection box.A test data set was constructed from multiple open source remote sensing image data sets,and the improved k-means++algorithm was used for anchor box clustering.Experimental results show that the model in this paper is significantly better than the comparison model in terms of mean average precision.At the same time,it can carry out real-time detection on low-computing power hardware and has practical application value.
作者
王磊
WANG Lei(Jiaxing Xiuzhou District Galaxy Surveying and Mapping Co.,Ltd.,Jiaxing 314031,China)
出处
《测绘与空间地理信息》
2023年第12期120-123,共4页
Geomatics & Spatial Information Technology
关键词
遥感影像
目标检测
轻量级网络
深度可分离卷积
非极大值抑制算法
remote sensing images
target detection
lightweight network
depth-separable convolution
non-maximum suppression algorithm
作者简介
王磊(1982-),男,江苏泗阳人,工程师,本科学历,主要从事工程测量、不动产测量等工作。
