False positive and false negative elimination technology based on thinking chain enhancement for power inspection

Read full text Summarise with VIRAT
Authors: Ying Li, Xiya Li, Zhuyi Rao
Cited by: 0
References: 8

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Abstract This paper aims to address the issues of false positives and false negatives in power inspections, which are caused by complex background interference, small-scale defects, and intra-class morphological differences. By constructing a multi-step reasoning discrimination framework based on thinking chain enhancement, the defect identification task is broken down into four steps: visual localization, attribute observation, rule comparison, and integrated decision-making. A hybrid model architecture combining BLIP-2 and GPT-4 is adopted, integrating two mechanisms: context-aware description generation and multi-round questioning validation, to enhance adaptability to complex scenarios. Experimental results show that the proposed method reduces the false positive rate to 0.062 and the false negative rate to 0.085 on a difficult sample test set, while the F1 score increases to 0.928, significantly outperforming traditional methods like YOLOv7 and DETR. Additionally, recall rates exceed 0.95 for typical defects, such as insulator flashover, demonstrating excellent detection sensitivity and task difficulty perception. This study verifies the effectiveness of thinking chain reasoning in power inspection and provides an interpretable and highly reliable solution for defect detection in complex industrial scenarios.

Fields of Study & Subjects

Power Line Inspection Robots Industrial Vision Systems and Defect Detection Infrastructure Maintenance and Monitoring

Publication details

Journal / source
Journal of Physics Conference Series
Volume
3192
Issue
1
Pages
012031–012031
Publication date
1 March 2026
Language
EN