基于改进Mask R-CNN的低空遥感实例分割算法  

Low-altitude remote sensing instance segmentation algorithm based on enhanced Mask R-CNN

作　　者：李冰锋[1,2] 王光耀 崔立志 LI Bingfeng;WANG Guangyao;CUI Lizhi(School of Electrical Engineering and Automation,Henan Polytechnic University,Jiaozuo 454000,China;Henan Key Laboratory of Intelligent Detection and Control of Coal Mine Equipment,Jiaozuo 454003,China)

机构地区：[1]河南理工大学电气工程与自动化学院,河南焦作454000 [2]河南省煤矿装备智能检测与控制重点实验室,河南焦作454003

出　　处：《兵器装备工程学报》2025年第2期168-176,共9页Journal of Ordnance Equipment Engineering

摘　　要：针对遥感领域图像目标繁杂、检测和分割精度不高的问题,提出一种改进Mask R-CNN算法。设计PMResNet-50结构作为主干网络,其中金字塔挤压注意模块可以促进局部和全局通道注意之间的信息交互作用,多层次特征聚合模块可以提高PMResNet-50对输入通道语义信息的高效聚合作用。在RoI Align前引入自校准卷积模块来扩大卷积层的感受野大小并对边界框和掩码框执行校准操作。在分割分支使用掩码预测平衡损失函数,对每个类别的正负样本梯度进行平衡,实现对损失梯度的平滑降低处理。在自建低空遥感数据集和iSAID-Reduce100数据集上进行测试,实验结果表明:改进后的算法在自建数据集上box AP和mask AP分别提升17.9%和15.0%,在iSAID-Reduce100数据集上box AP和mask AP达到49.62%和50.27%,该算法很好地完成了对遥感目标的检测和分割。In response to the challenges of complex object detection and low precision in image segmentation in the remote sensing domain,an enhanced version of the Mask R-CNN algorithm was proposed.The PMResNet-50 architecture was designed as the backbone network,with the Pyramid Squeeze Attention Modules facilitating information interaction between local and global channel attentions.Additionally,the Multilevel Feature Aggregation Module was employed to enhance the efficient aggregation of semantic information across input channels within PMResNet-50.A self-calibrating convolutional module was introduced before RoI Align to expand the receptive field size of convolutional layers and perform calibration operations on bounding boxes and mask boxes.In the segmentation branch,the Mask Prediction Balanced Loss function was utilized to balance gradients of positive and negative samples for each class,achieving a smooth reduction in loss gradient handling.Upon testing on both our self-built low-altitude remote sensing dataset and the iSAID-Reduce100 dataset,experimental results demonstrate that the improved algorithm achieves a 17.9%and 15.0%increase in box AP and mask AP,respectively,on our self-built dataset.Similarly,on the iSAID-Reduce100 dataset,the box AP and mask AP reach 49.62%and 50.27%,respectively.This algorithm effectively accomplishes detection and segmentation of remote sensing objects.

关 键 词：深度学习 图像处理 遥感图像 实例分割 改进Mask R-CNN算法 ResNet-50 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]