月表落石检测方法研究——基于梯度注入策略与结构驱动评价机制  

Research on detection method of lunar rockfall:based on gradient-injected strategy and structure-driven evaluation mechanism

作　　者：胡云泽 李杰[1] 何江[1] 袁强强[1] 郑莉[1] HU Yunze;LI Jie;HE Jiang;YUAN Qiangqiang;ZHENG Li(School of Geodesy and Geomatics,Wuhan University,Wuhan 430079,China)

机构地区：[1]武汉大学测绘学院,武汉430079

出　　处：《测绘工程》2024年第1期6-14,共9页Engineering of Surveying and Mapping

基　　金：国家自然科学基金资助项目(62071341);国家优秀青年科学基金资助项目(41922008)。

摘　　要：月表落石是了解月球演化进程的重要手段。由于人工识别落石特征既耗时又困难,所以落石检测逐渐倾向于利用深度学习手段。基于此,文中提出一种结合Transformer和梯度算子的月表落石有效检测方法。受落石结构的启发,设计一种基于落石成功检测数目的新评价机制。将3种目标检测算法和3种样本设置策略相结合,构建9个落石检测器。其中,样本设置通过调整样本比例和注入梯度信息以实现。实验结果表明,相比于现有基于CNN的落石检测方法,文中提出的月表落石检测方法在月表落石数据集RMaM-2020精度方面表现更优。同时,文中所提出的落石检测评价机制能够实现量化指标和可视化结果相统一。Lunar rockfall analysis is an important way to understand lunar crustal processes.Some researchers tend to detect rockfalls automatically by using deep learning because the manual mapping of rockfall features is time consuming and difficult.In this paper,we propose an effective lunar rockfall detection method that combine Transformer with gradient operators.Inspired by the rockfall structure,we design a new evaluation mechanism based on the number of successfully detected rockfalls.We construct nine rockfall detectors by combining three object detection algorithms with three sample setting strategies,which are achieved by adjusting the sample proportion and adding gradient operators.Compared with existing CNN-based rockfall detection methods,the experimental result shows that the proposed method performs better on lunar rockfall dataset,RMaM-2020,especially in terms of precision.Meanwhile,the proposed evaluation mechanism can keep the quantitative index consistent with the visualization results.

关 键 词：落石检测 深度学习 梯度算子 结构驱动 评价机制 

分 类 号：P407[天文地球—大气科学及气象学]