融合特征的深度学习遥感图像目标检测模型  

Fusion Features Based Deep Learning Remote Sensing Image Target Detection Model

作　　者：张永福[1] 宋海林 班越 汪西莉[1] ZHANG Yong-fu;SONG Hai-lin;BAN Yue;WANG Xi-li(School of Computer Science,Shaanxi Normal University,Xi’an 710119,China)

机构地区：[1]陕西师范大学计算机科学学院,陕西西安710119

出　　处：《计算机技术与发展》2021年第9期48-54,共7页Computer Technology and Development

基　　金：科技部第二次青藏高原综合科学考察研究项目(2019QZKK0405)。

摘　　要：为了提高目标检测模型对遥感图像中排列密集、尺度不一的目标,特别是小目标的检测性能,提出了融合特征的深度学习遥感图像目标检测模型和方法。模型采用小规模的网络结构,以应对标记样本较少的情况,并提出了融合多级特征的策略获取更为有效的特征,使模型在不增加检测时间的同时,提高遥感图像中较为密集且大小不一的目标的检测精度。模型中提出了一种新的后处理算法——分组融合剔除检测框算法,在剔除冗余检测框的同时微调检测框位置,使检测框对目标定位更精确,进一步提升检测精度。实验结果表明,所提模型在UCAS-AOD和RSOD-Dataset数据集上检测飞机,精度比Faster R-CNN的结果提高了4.2%和7.3%,漏检率和误检率均有降低。在UCAS-AOD数据集上检测更小的汽车目标,所提模型比Faster R-CNN检测精度提高了7.9%,漏检率下降了5.91%,误检率下降了2.06%。和Faster R-CNN相比,所提的融合处理和检测框后处理算法使得模型针对复杂场景中多尺度密集目标和小目标取得了更高的检测性能。In order to improve the performance of the object detection model for dense,different sizes targets,especially for small targets,in remote sensing images,a fusion features based deep learning remote sensing image target detection model is proposed.The model adopts small-scale network structure aiming at the small labeled sample data sets.The strategy of multi-level feature fusion is proposed to obtain more effective features,so that the model can improve the detection accuracy for dense and different sizes targets in remote sensing images without increasing the detection time.A new post-processing algorithm,namely packet fusion reject detection bounding boxes algorithm,is proposed,which can remove the redundant detection bounding box and fine-tune the position of the detection box,thus improving the box’s locating accuracy and further improving the detection accuracy.The experimental results on the UCAS-AOD and RSOD-Dataset remote sensing datasets for detecting aircraft and vehicle targets show that the proposed model achieves high detection performance for multi-scale and dense targets on small sample datasets.Compared with Faster R-CNN,it improves the detection accuracy by 4.2%and 7.3%for plane target,and by 7.9%for car targets,and reduces the missing ratio by 5.91%and false alarm rate by 2.06%.Compared with the existing models for dense and different sizes targets detection in complex remote sensing image,the proposed model achieves higher performance due to the feature fusion processing and the detection post-processing algorithm.

关 键 词：遥感图像 目标检测 卷积神经网络 特征融合 后处理 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]