摘要
光伏产业的面积统计是一个具有挑战性的问题,通过深度学习自动获取光伏面板的面积是一个可行的方案.为了快速准确地获取光伏面板的分布情况,本文设计了一个卷积神经网络模型用于提取遥感图像中的光伏面板.首先,以Res Net作为主干网络,结合金字塔池化模块构建了一个多尺度模型,以此表达各种尺度的光伏面板的视觉特征.然后,引入了非局部操作,融合长距离的上下文依赖关系,利用目标之间空间上的相关性更准确的提取前景目标.最后,提出了一种自适应上采样的方法,通过高分辨率输入图像的结构信息自适应地生成采样系数,指导低分辨率的语义特征图进行上采样,以此降低目标边缘模糊的问题.相比于三种较新的算法,在loU,精度和指标上F-Measure,本文的算法取得了最好的结果.
The statistics and management of the photovoltaic industry is a challenging problem,and it is a feasible solution to automatically obtain the area of photovoltaic panels by deep learning.In order to obtain the distribution of photovoltaic panels accurately and quickly a convolutional neural network model is designed to segment the photovoltaic panels in remote sensing images.First,using ResNet as the backbone network and combined with the pyramid pooling module,a multi-scale model is proposed to segment photovoltaic panels.Then,a non-local operation is introduced to fuse long-distance context dependencies to extract foreground objects more accurately.Finally,an adaptive upsampling module is proposed to upsample the low-resolution semantic feature map,and reducing the blurring of objects by using the structural information of high-resolution input image.Compared with the three new algorithms,the proposed model achieved the best results in terms of IoU,accuracy and F-Measure.
作者
宋灿
吴谨
朱磊
邓慧萍
SONG Can;WU Jin;ZHU Lei;DENG Hui-ping(College of Information Science and Technology,Wuhan University of Science and Technology,Wuhan 430081,China;Engineering Research Center of Metallurgical Automation and Measurement Technology,Ministry of Education,Wuhan 430000,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2021年第7期1485-1491,共7页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(61502358,61502357)资助。
关键词
光伏面板
遥感图像
深度学习
非局部操作
自适应上采样
photovoltaic panel
remote sensing
deep learning
non-local operation
adaptive upsample
