| With the development of ground observation technology on satellite remote sensing,the spatial resolution of the remote sensing data is getting higher and higher,and the data volume is also growing explosively.The problem of large distance between intraclass samples and consistent among inter-class samples in high-resolution remote sensing images bring great challenges to the semantic segmentation task of remote sensing images.Although the traditional combination of deep convolutional feature extraction network and the attention mechanism can extract the semantic feature information and improve the segmentation accuracy of the network,it also brings a large number of parameters and cause difficulties to achieve lightweight models.The existing methods of reducing the number of parameters to realize lightweight network mainly start from two perspectives: one is to use the existing lightweight network for mobile terminals as the feature extraction part of the model,and then improve the feature extraction ability of the network combined with other methods;the other is to design the lightweight network for feature extraction independently.Although these existing methods can greatly compress the parameters of the model,they often lead to a significant sacrifice of the segmentation performance.How to balance the segmentation performance of the network and the complexity of the model（from the perspective of parameters）has become the research focus.To solve the above problems,we design a lightweight attention-enhanced network that can maintain the satisfied segmentation performance and greatly reduce the number of parameters simultaneously.In this model,a multi-scale feature fusion lightweight module integrates different layers of convolutional neural networks（CNN）extracted from pre-trained deep neural networks.This module can not only achieve feature fusion,but also can greatly reduce the number of parameters of the model.Moreover,this paper introduces positional information in the self-attention mechanism to compensate for the loss of positional information caused by images in deep CNN.Finally,the introduction of covariance information is able to partially correct the generated attention map and improve the expression power of the attention mechanism.In order to verify the validity of the model,the experiments are conducted on the Vaihingen and Potsdam datasets published by the International Society of Photogrammetry and Remote Sensing（ISPRS）.Our model has achieved satisfactory results. |
