高分辨率遥感图像较大,图像上的目标尺度差异较大,对小目标识别效果差,并且分割过程中采样会造成信息丢失.针对上述问题,本文提出一种模块分组non local注意力机制的多级特征融合网络.以高分辨率神经网络(high-resolution network,HRNet)为骨架网络提取特征,并将多层特征融合网络(multi-level feature fusion network,MN)与分组non local注意力结合起来,利用反卷积和双线性插值组合的方法进行上采样,获得不同类型的图像的特征.在Vai‑hingen数据集上进行实验,分割精度达到91.46％,相比于其他分割网络,如Unet、Deeplabv3Plus等,该网络在一定程度上提高了遥感图像分割的识别效果.同时在WHU Building Dataset上验证了其对小目标的提取能力.
A Non-local Multi-layer Feature Fusion Method for Remote Sensing Image Semantic Segmentation Based on Module Grouping
The high-resolution remote sensing images are large,and the target scale differences on the image are large,but the recognition effect of small target is poor. In addition,information will be lost due to downsampling in the segmenta-tion process. To solve the above problems,this paper propos-es a multi-stage feature fusion network with module grouping and non-local attention mechanism. Firstly,HRNet was used as the skeleton Network to extract features,and multi-level feature fusion network (MN) was combined with grouped non-local attention. Deconvolution and bilinear interpolation were used for up-sampling. Get the features of different types of images. Finally,experiments were carried out on Vaihin-gen data set,and the segmentation accuracy reached 91.46％. Compared with other segmentation networks,such as Unet and Deeplabv3Plus,the recognition effect of remote sensing image segmentation was improved to a certain extent. At the same time,the small target extraction ability is verified in the WHU Building Dataset.
deep learningremote sensing imagesemantic segmentationnon localmulti-stage characteristics
NETL
NSTL
万方数据
