基于增强DeepLabV3网络的高分辨率遥感影像分类

High-resolution remote sensing image classification based on improved DeepLabV3 network

由于高分影像具有地物细节丰富、类别差异大等特点,现有的卷积神经网络影像分类方法普遍存在分类精度低、地物边界不准确等问题。鉴于此,本文提出一种基于增强DeepLabV3网络的影像分类模型。首先构建R-MCN网络结构,利用大小不同的卷积核并结合残差网络的思想进一步提取浅层网络的多尺度、多层次的特征信息;然后采用可学习的上采样方式,并将R-MCN提取的特征与高阶的语义信息相融合;最后通过提出的Mloss损失函数,获得遥感影像的地物分类结果。试验结果表明,相对于传统的卷积神经网络,本文方法能细化地物的边缘信息,改善分类效果,获得更高的影像分类精度。

Since high-resolution images usually have features such as rich details and large category differences,current methods for remote sensing classification of convolutional neural network generally perform poorly in classification accuracy and object boundary detection.In this paper,we propose an image classification model based on enhanced DeepLabV3 network.Firstly,R-MCN is built by combing residual networks and convolution kernels of different sizes,which is used to extract multi-scale,multi-level feature information of shallow networks.Then a learnable upsampling method is used to restore image size,and fuse the features extracted by R-MCN with high-level semantic information.Finally,a loss function named Mloss is built to achieve classification results of remote sensing images.The experimental results demonstrate that the proposed method can refine object boundaries,improve classification performance,and obtain higher accuracy of image classification compared with traditional convolutional neural network.