基于改进特征金字塔的遥感图像全景分割  

Panoptic Segmentation of Remote Sensing Image Based on Improved Feature Pyramid Network

作　　者：幸羚 朱雯婕 徐昇[1] XING Ling;ZHU Wenjie;XU Sheng(College of Computer Science and Technology,Nanjing Forestry University,Nanjing 210037)

机构地区：[1]南京林业大学信息科学技术学院,南京210037

出　　处：《计算机与数字工程》2023年第4期893-898,905,共7页Computer & Digital Engineering

基　　金：国家自然科学基金青年科学基金项目(编号:62102184);江苏省自然科学基金青年科学基金项目(编号:BK20200784);中国博士后科学基金面上项目(编号:2019M661852);江苏省高等学校大学生创新创业训练计划项目“针对遥感图像的改进生成对抗网络超分辨率重建”(编号:202210298042Z)资助。

摘　　要：把全景分割技术应用于蕴含着复杂地物信息的遥感图像更能满足实际需求,但迄今全景分割模型大多采用不同的方式表征背景未定形区和前景实例对象,利用两个独立的网络分别实现语义和实例分割任务,导致需要融合操作。全景特征金字塔网络通过单一网络实现了联合任务,简化复杂操作,但其对实例对象中的小目标分割效果不够理想,边缘信息比较模糊。论文针对以上问题进行改进。首先,在不同大小的残差网络(ResNet50、ResNet101)上添加特征金字塔网络,改善多尺度的特征提取。然后在上采样时利用双线性插值法使得边缘信息更加清晰。最后,通过调整损失函数的权重,重新加权,得到具有更高全景质量的两个模型R50-FPN、R101-FPN。在COCO数据集上与全景质量为41.3%的OANet模型相比,模型R50-FPN、R101-FPN全景质量分别提高了0.2%和1.7%。此外,分别比较两模型在遥感图像和街景图像上的全景分割表现,实验表明两个模型在街景图像上的分割精度都更高。而模型R101-FPN相比于R50-FPN在边缘处理上更加精准,全景质量提高了1.5%。The application of panoptic segmentation technology to remote sensing images containing complex ground object information can better meet the actual needs.But so far,most panoptic segmentation models use different ways to represent background unshaped area and foreground instance objects,and use two independent networks to achieve semantics and segmentation tasks respectively,resulting in the need for fusion operations.Panoramic feature pyramid network realizes joint tasks through a single network simplifying complex operations.However,its segmentation effect on small objects in instance objects is not ideal,and the edge information is relatively blurred.This paper improves on the above problems.Firstly,feature pyramid networks are added to residual networks of different sizes(ResNet50,ResNet101)to improve multi-scale feature extraction.Then through using the bilinear interpolation method during upsampling,the edge information becomes clearer during upsampling.Finally,by adjusting the weight of the loss function and re-weighting,two models R50-FPN and R101-FPN with higher panoramic quality are obtained.Compared with the OANet model with a panorama quality of 41.3%on the COCO dataset,the panorama quality of R50-FPN and R101-FPN is improved by 0.2%and 1.7%respectively.In addition,comparing the panoptic segmentation effect of the two models on remote sensing images and street view images respectively,experiments show that both models have higher segmentation accuracy on street view images.Besides,compared with R50-FPN,R101-FPN is more accurate in edge processing,the panoptic quality is improved by 1.5%.

关 键 词：全景分割 遥感图像 特征金字塔网络 双线性插值 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]