软件导刊2025，Vol.24Issue(6)：151-159,9.DOI:10.11907/rjdk.241162
基于对比学习的遥感图像小样本目标检测算法
Few-Shot Object Detection Algorithm in Remote Sensing Images Based on Contrastive Learning
摘要
Abstract
Aiming at the problem that the current remote sensing image object detection algorithm relies on large-scale labeled datasets to per-form well,and the accuracy rate is low when detecting novel classes,a new remote sensing image few-shot object detection model is proposed based on the classic detection model Faster-RCNN.First,the model effectively alleviates overfitting on balanced sub-datasets by introducing meta-training methods.Second,it utilizes contrastive learning in the hyperspace to assist the model in better encoding novel class objects and base class objects,thereby achieving more accurate object classification.This approach improves the detection performance on novel class ob-jects while mitigating the catastrophic forgetting caused by the introduction of novel classes.In addition,a non-maximal suppression algorithm suitable for few-shot object detection scenario is proposed to effectively solve the repeated detection of the same object.Experimental results on the large-scale benchmark dataset DIOR show that,compared with the suboptimal network MM-RCNN,the proposed method improves the accuracy of new class by 3.5%,base class by 8.7%,and all classes by 6.4%.At the same time,the effectiveness of the method is further proved by NWPU VHR-10 dataset.关键词
目标检测/对比学习/非极大抑制/小样本/遥感图像Key words
object detection/contrastive learning/non-maximal suppression/few-shot/remote sensing image分类
信息技术与安全科学引用本文复制引用
李宗霖,符颖..基于对比学习的遥感图像小样本目标检测算法[J].软件导刊,2025,24(6):151-159,9.基金项目
四川省科技厅重大科技专项(2019ZDZX0005) （2019ZDZX0005）
