基于改进DeeplabV3+的农村遥感图像屋顶分割

SOLAR ROOFTOP SEGMENTATION IN RURAL AREAS BASED ON IMPROVED DEEPLABV3+

现有的太阳能屋顶分割方案大多数针对城市地区,而农村地区的遥感图像不清晰、房屋数量多且呈小区域分散的状态,使得现有模型对农村地区屋顶分割效果不佳。因此,提出基于改进DeeplabV3+的农村地区遥感图像屋顶分割算法。改进的算法在原DeeplabV3+的基础上添加了棋盘格平滑模块,以缓解空洞卷积引起的棋盘格效应;为了进一步恢复屋顶的边缘细节信息,提出融合多低水平特征和通道注意力模块。实验结果表明,提出的算法在农村地区遥感图像屋顶分割任务中取得了良好的效果,模型在验证集上的像素准确率约达到了96.05%,前景的交并比约达到了91.62%,指标与原DeeplabV3+模型相比分别提升了0.022 1和0.017 4,表明提出的方法更加适用于农村地区遥感图像的屋顶分割。

Most of the existing solar rooftop segmentation schemes are aimed at urban areas, while remote sensing images in rural areas are not clear, there are a large number of houses and they are scattered in small areas, making the existing model poor in roof segmentation in rural areas. Therefore, an improved remote sensing image rooftop segmentation algorithm for rural areas based on improved DeeplabV3+ is proposed. The improved algorithm added a checkerboard smoothing module to the original DeeplabV3+ to alleviate the checkerboard effect caused by atrous convolution. In order to further restore the rooftop edge detail information, a multi-low level feature and channel attention module were proposed. The experimental results show that the proposed algorithm has achieved good results in the rooftop segmentation task of remote sensing images in rural areas. The pixel accuracy of the model on the validation set has reached 96.05%, and the prospect ratio has reached 91.62%, which are improved by 0.022 1 and 0.017 4 respectively compared with the original DeeplabV3+, indicating that the proposed method is more suitable for roof segmentation of remote sensing images in rural areas.