不同特征提取策略的场景变化检测性能评估

Performance evaluation of different feature extraction strategies for high-resolution remote sensing image scene change detection

分类后比较是场景变化检测的常用方法,其中分类结果直接决定了变化检测的精度,而图像特征的选择和提取是实现高精度分类的关键。为评估不同特征提取策略对变化检测精度的影响,本文选取预训练的艾历克斯网络(Alex Net)、视觉几何图形小组十六层网络(VGG16)、视觉几何图形小组十九层网络(VGG19)、谷歌模块组装型网络(Goog Le Net)、十八层残差网络(Res Net18)、五十层残差网络(Res Net50)、压缩型网络(Squeeze Net)和十九层黑暗网络(Dark Net19)等网络提取卷积神经网络(CNN)特征,以及传统的纹理特征和颜色特征,对高分辨率遥感影像进行场景变化检测。结果表明:Squeeze Net、Dark Net19和Res Net50在场景分类中表现最出色,三者变化检测效果亦最佳,总体精度均高达0.95,KAPPA系数达0.90。而传统特征的变化检测精度较CNN特征中表现最差的VGG19还要逊色,验证了CNN特征对场景变化检测的优越性能。

Comparison after classification is a commonly used method for scene change detection,and the classification results directly determine change detection accuracy,while features selection and extraction are the key to achieve classification with high accuracy.To evaluate the influence of different feature extraction strategies on the accuracy of change detection,eight pre-trained networks including AlexNet,visual geometry group16(VGG16),visual geometry group19(VGG19)、Google inception net(GoogLeNet),residual network18(ResNet18),residual network50(ResNet50),SqueezeNet,DarkNet19 were selected to extract convolutional neural networks(CNN)features,as well as conventional texture and color features to conduct change detection for for high-resolution remote sensing images.The results showed that SqueezeNet,DarkNet19 and ResNet50 performed the best in scene classification and also had the best change detection accuracy,with the overall accuracy as high as 0.95 and the KAPPA coefficient of 0.90.However,the accuracy of traditional features was even worse than VGG19,which had the worst performance among CNN features,indicating the superior performance of CNN features for scene change detection.