A Self-Supervised Enhancement Method on Large Foundation Vision Model for Power System Detection

Le résumé fourni par la source

Nowadays, the development of power systems is trending towards becoming digital and intelligent. Recently, deep learning methods have been increasingly applied to the inspection work of power system operation and maintenance. However, the application scenarios are complex, which involve different data distribution gaps. In the industrial production environment, a single model typically serves a single task. To address the problem, we proposed DINOv2-based knowledge-projection supplementary Network (DKSN), which utilizes the remarkable unified encoding capability of large foundation vision model to acquire general feature representation. These representations are optimized by our proposed knowledge projection model and self-supervised detail recovery model, which provides sufficient prior knowledge for subsequent decoder to generate the final detection result. DKSN has demonstrated strong domain adaptation capability in detection tasks across multiple scenarios, achieving a recall of$\mathbf{9 9. 1 5 \%}$and a mAP of$\mathbf{9 5. 1 8 \%}$on the final test dataset.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.