To address the challenge that remote sensing image object detection heavily depends on extensive annotated data and that existing efficient fine-tuning methods for general parameters exhibit insufficient spatial perception when adapted to remote sensing tasks, a vision-language collaborative efficient fine-tuning method for multimodal large models is proposed. Firstly, the existing dataset is reconstructed to construct instruction data suitable for large language and vision assistant (LLaVA) fine-tuning. Secondly, combining the multi-cognitive visual adapter (Mona) and the quantized low-rank adaptation (QLoRA) method, the LLaVA model is efficiently fine-tuned on the instruction fine-tuning dataset. Finally, the target category and coordinate information are obtained in the inference stage. Experiments on the NWPU VHR-10 dataset show that compared with the pre-trained LLaVA model, the collaborative fine-tuning model has greatly improved in F1 score, accuracy, and mIoU. Its mAPs on both NWPU VHR-10 and DIOR datasets are comparable to those of traditional detection models. This method still performs well in low-annotation scenarios, providing a new technical approach for the efficient application of multimodal large models in remote sensing vision tasks.
近年来,以GPT-4和大型语言和视觉助手(Large Language and Vision Assistant, LLaVA)[12]为代表的多模态大语言模型(Large Language Model, LLM)在视觉—语言联合理解与推理方面展现出强大能力。这类模型通过在超大规模图像—文本对上进行预训练,构建通用的视觉概念与语义知识库,并展现出强大的零样本迁移能力。这为破解遥感目标检测的数据瓶颈提供了全新的技术路径。然而,如何将此类通用大模型高效适配至遥感视觉任务,仍面临核心挑战。现有的参数高效微调技术主要针对语言模型设计,通过低秩矩阵模拟参数更新[13-15],或者固定预训练主干网络的原始参数,并添加额外可训练结构[16-19],虽然能高效调整语义理解能力,却难以优化模型对图像空间结构、多尺度目标及复杂背景的理解能力,导致其直接应用于遥感图像目标检测时精度有限。因此,探索一种能够同时增强视觉空间感知与语言语义理解的高效微调框架,成为推动多模态大模型在遥感领域落地的关键。
实验结果采用目标检测常用的评价指标:精确率(Precision)、召回率(Recall)、F1值(F1 Score)、准确率(Accuracy)、平均交并比(mean Intersection over Union, mIoU)和平均精度均值(mean Average Precision, mAP)。针对算法效率,采用训练参数量、训练参数占比、训练峰值显存和训练耗时作为评价指标。
DALALN, TRIGGSB. Histograms of oriented gradients for human detection[C]//Proceedings of the 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. Piscataway, USA: IEEE, 2005:886-893.
[2]
FREUNDY, SCHAPIRER E. Experiments with a new boosting algorithm[C]//Proceedings of the Thirteenth International Conference on International Conference on Machine Learning. New York, USA: ACM, 1996:148-156.
[3]
VAPNIKV N. The nature of statistical learning theory[M]. New York, USA: Springer, 1995:1-203.
LIUW, ANGUELOVD, ERHAND, et al. SSD: single shot multibox detector[M]//LEIBE B, MATAS J, SEBE N, et al. Computer vision-ECCV 2016. Cham, Swizterland: Springer, 2016:21-37.
[6]
REDMONJ, DIVVALAS, GIRSHICKR, et al. You only look once: unified, real-time object detection[C]//Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway, USA: IEEE, 2016:779-788.
ZHAOY A, LVW Y, XUS L, et al. DETRs beat YOLOs on real-time object detection[C]//Proceedings of the 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway, USA: IEEE, 2024:16965-16974.
[9]
VASWANIA, SHAZEERN, PARMARN, et al. Attention is all you need[J]. Advances in Neural Information Processing Systems, 2017,30:5998-6008.
[10]
XUD Q, WUY Q. FE-YOLO: a feature enhancement network for remote sensing target detection[J]. Remote Sensing, 2021,13(7):No.1311.
[11]
LIY S, LUOJ W, ZHANGY J, et al. Learning to holistically detect bridges from large-size VHR remote sensing imagery[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024,46(12):11507-11523.
[12]
LIUH T, LIC Y, WUQ Y, et al. Visual instruction tuning[J]. Advances in Neural Information Processing Systems, 2023,36:34892-34916.
[13]
ZHONGH Q, SHENS C, CAIK, et al. Serial low-rank adaptation of vision transformer[C]//Proceedings of the 2025 IEEE International Conference on Multimedia and Expo. Piscataway, USA: IEEE, 2025. DOI: 10.1109/ICME59968.2025.11209194 .
[14]
LEIC, LIA, YAOH, et al. Rethinking token reduction with parameter-efficient fine-tuning in ViT for pixel-level tasks[C]//Proceedings of the 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway, USA: IEEE, 2025:14954-14964.
ZHUB E, NIUY L, HANY C, et al. Prompt-aligned gradient for prompt tuning[C]//Proceedings of the 2023 IEEE/CVF International Conference on Computer Vision. Piscataway, USA: IEEE, 2023:15613-15623.
[17]
HEX H, LIC Y, ZHANGP C, et al. Parameter-efficient model adaptation for vision transformers[DB/OL]. (2023-07-13)[2026-01-25].
[18]
YIND S, YANGY R, WANGZ C, et al. 1% VS 100%: parameter-efficient low rank adapter for dense predictions[C]//Proceedings of the 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway, USA: IEEE, 2023:20116-20126.
[19]
GONGM, DENGY N, QIN A, et al. Structure-learnable adapter fine-tuning for parameter-efficient large language models[C]//Proceedings of the International Conference on Artificial Intelligence and Computational Engineering 2025. London, UK: IET, 2025:225-230.
[20]
YIND S, HUL Y, LIB, et al. 5%>100%: breaking performance shackles of full fine-tuning on visual recognition tasks[C]//Proceedings of the 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway, USA: IEEE, 2025:20071-20081.
[21]
DETTMERST, HOLTZMANA, PAGNONIA, et al. QLoRA: efficient finetuning of quantized LLMs[J]. Advances in Neural Information Processing Systems, 2023,36:10088-10115.
[22]
HOULSBYN, GIURGIUA, JASTRZEBSKIS, et al. Parameter-efficient transfer learning for NLP[C]//Proceedings of the 36th International Conference on Machine Learning. New York, USA: PMLR, 2019:2790-2799.
[23]
HUE J, SHENY L, WALLISP, et al. LoRA: low-rank adaptation of large language models[DB/OL]. (2021-10-16)[2026-01-25].