Article Details

Abstract

To address the challenges of insufficient accuracy and generalization in image recognition for small-sample scenarios in intelligent power inspection, this study proposes a collaborative recognition method that integrates Multimodal Large Language Models (MLLMs) with the YOLOv8 algorithm. During the training phase, real images and prompts are used to guide MLLMs in generating multi-angle, multi-state augmented samples. These samples are employed to train the YOLOv8 model, thereby enhancing its generalization capability and recognition accuracy. In the inference stage, a partitioning method based on multi-level confidence scores from YOLOv8 is designed, which dynamically determines whether to incorporate the large model for joint reasoning according to predefined thresholds. Experimental results based on substation inspection data demonstrate that the proposed method effectively improves recognition performance under small-sample conditions, validating its feasibility and effectiveness.

Keywords

MLLMs; YOLOv8; intelligent inspection; large-small model collaboration; F1-score

References

1. [1] National Energy Administration Guidelines for Intelligent Inspection Technology of Wind Farm: NB/T10596-2021 [S]. 2021
2. [2] Lin Yuanmin, Wang Xin, Huang Jinzhao, etc Real time fault image recognition system for intelligent distribution network based on artificial neural network [J]. Power Grid and Clean Energy, 2024, 40 (6): 130-136
3. [3] Redmon J，Divvala S，Girshick R，et al. You Only Look Once: Unified, Real-Time Object Detection [C]//2016 IEEE Conference on Computer Vision and Pattern Recognition （CVPR），2016：779-788.
4. [4] Chen Furong, Xiao Mingming Research on Improved Small Object Detection Algorithm Based on YOLOv5 [J]. Modern Information Technology, 2023, 7 (3): 55-60, 65
5. [5] Chen Long, Zhang Jianlin, Peng Hao, etc Multi scale attention and domain adaptive small sample image recognition [J]. Optoelectronic Engineering, 2023, 50 (4): 66-80
6. [6] Wang Hongjiang, Liu Jinsheng, Zhao Hong, etc Tree line grounding fault identification based on improved YOLOv8 model [J]. Journal of Liaoning University of Engineering and Technology (Natural Science Edition), 2025, 44 (1): 113-119
7. [7] Xia Qing Research and Implementation of UAV Target Detection and Tracking Based on YOLOv8 [D]. Yangzhou: Yangzhou University, 2025
8. [8] Li Gang, Fang Hong, Liu Yunpeng, etc Large Model Driven Technology in New Power Systems: Current Status, Opportunities, and Challenges [J]. High Voltage Technology, 2024, 50 (7): 2864-2878
9. [9] Pan Rujiang, Chen Jiongyi, Wang Jianbo, etc Intelligent modeling method for system architecture view based on big language model [J]. Systems Engineering and Electronic Technology, 2025, 47 (12): 3912-3923
10. [10] Tian Hongkun, Jiang Nan, Ren Tao UAV wire detection based on cross modal interaction fusion and global feature calibration [J]. Control and Decision, 2026, 41 (3): 639-650

About the authors

Chen Bei (1982-), female, Han ethnicity, born in Nanjing, Jiangsu, associate professor, master's degree, research direction: intelligent modeling, machine learning.