基于改进DeepLabV3+的遥感图像分割方法OA北大核心CSTPCD

Because of the high resolution of remote sensing images,convolutional layers need to enlarge their receptive fields to capture richer semantic information.In the process of remote sensing image segmentation,the larger dilation rate is adopted for the DeepLabV3+model to achieve a larger receptive field,leading to the issue of grid pseudo-artifacts.Therefore,an optimized DeepLabV3+model is proposed with improvements to address the problem of grid pseudo-artifacts.A smoothing grid pseudo-artifact module is introduced before the atrous spatial pyramid pooling(ASPP)to mitigate the impact of grid pseudo-artifacts on segmentation tasks.Subsequently,a pointwise convolution is added after each dilated convolution in the ASPP module to retain more spatial information.The activation function of dilated convolutions is replaced with LeakyReLU.The efficient channel attention(ECA)mechanism is introduced into DeepLabV3+.By validation on the GID15 and Postdam remote sensing datasets,the improved model demonstrates significant enhancements in terms of accuracy and mean intersection over union(MIoU)in comparison with the baseline DeepLabV3+model.This validates that the proposed network adjustments can effectively improve the accuracy of remote sensing image segmentation.