Project Details / 项目资讯
- Company / 公司:
- GitHub URL / GitHub 链接: JointNet-A-Common-Neural-Network-for-Road-and-Building-Extraction
Description / 描述
JointNet is a novel neural network which aims to achieve the following:
- Road and building extraction from optical satellite images via information extraction from linear shapes and large-scale objects
- Utilization of dense atrous convolution block that not only maintains the feature propagation efficiency of dense connectivity, but also to achieve a larger receptive field
- Utilization of focal loss function to imbalance problem between road centerline target and its background
- Replacement of Batch Normalization (BN) layer by Group Normalization (GN) layer to tackle small training batch sizes
JointNet 是一个崭新的神经网络,旨在实现以下目标:
- 通过从线性形状和大尺度对象中提取的信息进而从光学卫星图像中提取道路和建筑
- 利用稠密的空洞卷积块,不仅保持特征传播效率的稠密连接,而且还实现了更大的网络视野
- 使用焦点损失函数来解决道路中心线目标与其背景之间的不平衡的问题
- 替换批量归一化(BN)层为组归一化(GN)层,以应对小的训练批次
Reference
[1] Zhang Z, Wang Y. JointNet: A Common Neural Network for Road and Building Extraction. Remote Sensing. 2019; 11(6):696. https://doi.org/10.3390/rs11060696