遥感影像目标的尺度特征卷积神经网络识别法

Object detection in remote sensing imagery based on convolutional neural networks with suitable scale features

高分辨率遥感影像的目标检测与识别,是高分对地观测系统中影像信息自动提取及分析理解的重要内容。针对传统影像目标检测与识别算法中人工设计特征稳健性与普适性差的问题,本文提出基于高分辨率遥感影像目标尺度特征的卷积神经网络检测与识别方法。首先通过统计遥感影像目标的尺度范围,获得卷积神经网络训练与测试过程中目标感兴趣区域合适的尺度大小。然后根据目标感兴趣区域合适的尺度,提出基于高分辨率遥感影像目标尺度特征的卷积神经网络检测与识别架构。通过WHU-RSone数据集对本文卷积神经网络架构与Faster-RCNN架构对比测试验证。试验结果表明,本文架构ZF模型和本文架构VGG-16模型的mean average precision(mAP)分别比Faster-RCNNZF模型和Faster-RCNNVGG-16模型提高8.17%和8.31%,本文卷积神经网络架构可获得良好的影像目标检测与识别效果。

Object detection and recognition in high spatial resolution remote sensing images (HSRI) is an important part of image information automatic extraction, analysis and understanding in high resolution earth observation system. The robustness and universality of traditional object detection and recognition algorithms using artificial design object feature are poor. To solve these problems, object detection and recognition in HSRI based on convolutional neural networks (CNN) with suitable scale features is proposed. Firstly, the suitable scale of the region of interest (ROI) of object is obtained by statistic the scale range of object in HSRI in the process of training and testing of CNN. Then, a CNN framework for object detection and recognition in HSRI is designed according to the suitable object ROI scale. The mean average precision (mAP) of the proposed CNN framework and Faster-RCNN is tested using the WHU-RSone data set. The experimental results show that the mAP of ZF model and VGG-16 model of the proposed CNN framework are 8.17% and 8.31% higher than that of Faster R-CNN ZF model and Faster R-CNN VGG-16 model, respectively. The proposed CNN framework can obtain good object detection and recognition results.