【摘要】 针对高分辨率遥感图像建筑物分割方法存在分割完整度低和模型参数量大的问题，提出一种轻量化的编码器-解码器结构网络LED-Net (lightweight encoder decoder network)。编码器使用带有通道注意力机制的残差网络，使得网络充分提取图像的特征信息；解码器使用轻量化的数据依赖上采样模块，增加建筑物分割完整度，减少模型参数量。使用INRIA Aerial Image与Massachusetts Buildings数据集对该算法进行实验，LED-Net相较DeepLabV3+、Building-A-Net等模型，减少了模型参数量，提升了分割精确度及预测图中建筑物的整体和边缘分割完整度。更多还原

【Abstract】 Aiming at the problems of low segmentation completeness and the large number of model parameters in the building segmentation methods of high-resolution remote sensing images,a lightweight encoder-decoder structure network LED-Net(lightweight encoder decoder network)was proposed.The network encoder used a residual network with a channel attention mechanism to enable the network to fully extract image feature information.The decoder used a lightweight data-dependent upsampling module to increase the building segmentation completeness and reduce the number of model parameters.The algorithm was experimented with the INRIA Aerial Image and Massachusetts Buildings datasets.Results show that LED-Net reduces the number of model parameters and improves the segmentation accuracy and the completeness of the overall and edge segmentation of buildings in the prediction map compared to DeepLabV3+,Building-A-Net,and etc.更多还原