Class-Incremental Few-Shot Object Detection with Distillation Response in Remote Sensing Images

GUAN Wen-bin; WU Xiao-hong; CHEN Hong-gang; HE Xiao-hai

doi:10.3969/j.issn.2096-6091.2023.21.001

您当前的位置：

首页 >

文章列表页 >

Class-Incremental Few-Shot Object Detection with Distillation Response in Remote Sensing Images

RESEARCH PAPER | 更新时间：2025-03-26

- Class-Incremental Few-Shot Object Detection with Distillation Response in Remote Sensing Images
- New Generation of Information Technology Vol. 6, Issue 21, Pages: 1-7(2023)
- 作者机构：
  
  四川大学电子信息学院，四川成都 610065
- 作者简介：
- 基金信息：
- DOI：10.3969/j.issn.2096-6091.2023.21.001
  CLC： TP751.1
- Published：15 November 2023
- 稿件说明：
移动端阅览
官文彬, 吴晓红, 陈洪刚, 等. 基于蒸馏响应的类增量少样本遥感目标检测研究[J]. 新一代信息技术, 2023, 6(21): 01-07

GUAN Wen-bin, WU Xiao-hong, CHEN Hong-gang, et al. Class-Incremental Few-Shot Object Detection with Distillation Response in Remote Sensing Images[J]. New Generation of Information Technology, 2023, 6(21): 01-07
官文彬, 吴晓红, 陈洪刚, 等. 基于蒸馏响应的类增量少样本遥感目标检测研究[J]. 新一代信息技术, 2023, 6(21): 01-07 DOI： 10.3969/j.issn.2096-6091.2023.21.001.

GUAN Wen-bin, WU Xiao-hong, CHEN Hong-gang, et al. Class-Incremental Few-Shot Object Detection with Distillation Response in Remote Sensing Images[J]. New Generation of Information Technology, 2023, 6(21): 01-07 DOI： 10.3969/j.issn.2096-6091.2023.21.001.

摘要

近年来，少样本遥感目标检测越来越受到人们的关注，其目的是使目标检测器仅用少量标记的遥感图像样本就能检测到新的目标类别。但主流的少样本遥感目标检测器在学习新类知识的时候，必需要有基类样本的辅助。而在现实世界中，目标类别往往是不断变化的，且环境也在不断变换。这种目标动态性给传统的少样本目标检测方法带来了巨大的障碍，因为它们严重依赖基类样本来学习新类别，这就使这些少样本遥感目标检测器无法无限制的对新类别进行学习，而且随着学习的类别数不断增多，在内存使用量、存储用量和计算量上都将产生严重负担。本研究从实际角度出发，提出了基于蒸馏响应的类增量遥感少样本目标检测算法（Class-incremental Few-Shot Object Detection with Distillation Response in remote sensing images，DR-CFSOD），以实现在不需要基类数据辅助的情况下，利用新类数据独自完成学习任务。此外，本文所研究的算法为了缓解增量学习场景中经常发生的灾难性遗忘问题，通过提取从以前的类中学习到的知识，保留了模型的基类判别能力。为了验证所提方法的有效性，在两个不同的遥感图像数据集上进行了广泛的实验，结果表明，DR-CFSOD比基线方法具有更好的性能。

Abstract

In recent years

few-shot object detection (FSOD) in remote sensing images (RSIs) has garnered significant attention from the research community. This technique aims to empower object detectors to identify novel object categories using a minimal number of labeled samples within RSIs. The importance of this research is underscored by the vast amounts of unlabeled remote sensing data available

making the efficient utilization of limited labeled data crucial. The core challenge in FSOD for RSIs lies in the requirement for base class samples during the learning process of novel class knowledge. In practical scenarios

object categories are constantly evolving and adapting to changing environments. This dynamic nature poses a significant obstacle for traditional FSOD methods

as they rely heavily on base class samples to learn novel class. However

such a dependency limits the ability of these methods to adapt tonovel class indefinitely. Furthermore

as the number of learned categories accumulates

the memory usage

storage requirements

and computational demands increase significantly. This scalability issue can become a major bottleneck

especially when dealing with large-scale remote sensing datasets. To address these challenges

we propose a class-incremental few-shot object detection method based on distillation response in remote sensing images (DR-CFSOD). This innovative approach differs significantly from existing methods by enabling the learning of novel class independently

without relying on base class. This crucial aspect alleviates the need for continuously incorporating base class samples

thereby overcoming the limitations of traditional FSOD methods. Moreover

DR-CFSOD addresses the catastrophic forgetting problem that often occurs in incremental learning scenarios. By distilling the knowledge learned from previous classes

the algorithm preserves the discriminative ability of the model

ensuring that it can accurately detect both base and novel class. To validate the effectiveness of our proposed method

we conducted extensive experiments on two diverse remote sensing image datasets. The results obtained demonstrate the superior performance of DR-CFSOD compared to baseline methods. Our method not only achieves competitive detection accuracy but also exhibits superior scalability and adaptability to new categories. Concurrently

our proposed method boasts an impressive 95FPS real-time detection performance

ensuring swift and accurate identification. This outstanding feature provides a solid foundation for seamless integration and effective application in practical scenarios

further enhancing its utility and value.

关键词

Keywords

references

SHMELKOV K , SCHMID C , ALAHARI K . Incremental learning of object detectors without catastrophic forgetting [C ] // 2017 IEEE International Conference on Computer Vision (ICCV) . Piscataway : IEEE , 2017 : 3400 - 3409 .

PENG C , ZHAO K , MAKSOUD S , et al . SID: Incremental learning for anchor-free object detection via selective and inter-related distillation [J ] . Computer Vision and Image Understanding , 2021 , 210 : 103229 .

LI D W , TASCI S , GHOSH S , et al . RILOD: Near real-time incremental learning for object detection at the edge [C ] // Proceedings of the 4th ACM/IEEE Symposium on Edge Computing . New York : ACM , 2019 : 113 - 126 .

康梦雪 , 张金鹏 , 马喆 . 等 . 基于高级语义特征蒸馏的增量式连续目标检测方法 [J ] . 现代防御技术 , 2024 , 52 ( 1 ): 41 - 48 .

刘冬梅 , 徐洋 , 吴泽彬 , 等 . 基于边框距离度量的增量目标检测方法 [J ] . 计算机科学 , 2022 , 49 ( 8 ): 136 - 142 .

PEREZ-RUA J M , ZHU X T , HOSPEDALES T M , et al . Incremental few-shot object detection [C ] // 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . Piscataway : IEEE , 2020 : 13846 - 13855 .

LI P , LI Y , CUI H , et al . Class-incremental few-shot object detection [EB/OL ] . ( 2021-12-28 )[ 2023-12-22 ] . https://arXiv.org/abs/ 2105.07637 https://arXiv.org/abs/2105.07637 .

DONG N , ZHANG Y Q , DING M L , et al . Incremental-DETR: Incremental few-shot object detection via self-supervised learning [J ] . Proceedings of the AAAI Conference on Artificial Intelligence , 2023 , 37 ( 1 ): 543 - 551 .

李莉萍 . 基于注意力机制和小样本增量学习的遥感影像目标检测 [D ] . 西安 : 西安电子科技大学 , 2023 .

LI L J , YAO X W , CHENG G , et al . Solo-to-collaborative dual-attention network for one-shot object detection in remote sensing images [J ] . IEEE Transactions on Geoscience and Remote Sensing , 2021 , 60 : 1 - 11 .

LI K , WAN G , CHENG G , et al . Object detection in optical remote sensing images: A survey and a new benchmark [J ] . ISPRS Journal of Photogrammetry and Remote Sensing , 2020 , 159 : 296 - 307 .

EVERINGHAM M , VAN GOOL L , WILLIAMS C K I , et al . The pascal visual object classes (VOC) challenge [J ] . International Journal of Computer Vision , 2010 , 88 ( 2 ): 303 - 338 .

WANG X , HUANG T E , DARRELL T , et al . Frustratingly simple few-shot object detection [C ] // Proceedings of the 37th International Conference on Machine Learning . Brookline : Microtome Publishing , 2020 : 9919 - 9928 .

ZHANG T Y , ZHANG X R , ZHU P , et al . Generalized few-shot object detection in remote sensing images [J ] . ISPRS Journal of Photogrammetry and Remote Sensing , 2023 , 195 : 353 - 364 .

FENG T , WANG M , YUAN H J . Overcoming catastrophic forgetting in incremental object detection via elastic response distillation [C ] // 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . Piscataway : IEEE , 2022 : 9427 - 9436 .

Views

334

下载量

CSCD

Alert me when the article has been cited

提交

Tools

Publicity Resources

No data

Related Author

No data

Related Institution

No data

⁰