面向遥感影像场景的深度卷积神经网络递归识别模型

Remote Sensing Image Scene Oriented Convolutional Neural Network Recursive Recognition Model

在遥感影像场景识别过程中,针对利用卷积神经网络进行固定格网影像场景识别时存在类间可分性不高和局部细节粗糙等问题,提出一种深度卷积神经网络递归识别模型(DCNN-RR)。该模型首先构建卷积层、采样层交替的多层卷积神经网络进行遥感影像多分辨率场景训练。然后,根据格网影像softmax概率计算场景类间混淆指数(Confusion Index,CI),四分格网递归进行卷积神经网络识别,并采用多重窗口滑动递归微调直至CI达到峰值来精准定位场景目标。通过高分辨遥感影像实验表明该模型可适应不同尺度地物的变化,相比固定格网影像显著提高了场景识别精度,局部细节也更为精细。

In order to solve low separability and rough details in scene recognition, remote sensing image scene oriented convolutional neural network recursive recognition model is presented.Firstly,deep convolu- tional neural network with multi-convolutional layers and multi-pooling layers is constructed by multi-res- olution scenes.Then quad-grids are subdivided to DCNN scene recursive recognition based on Confusion In- dex （CI）by softmax probability,and multi-sliding windows are used to tune recursively for accurately loca ting scene targets.Experimental results show that the proposed model can adapt scene recognition with dif- ferent scale,and significantly improve the accuracy compared with the commonly used DCNN.