A2MADA-YOLO: Attention Alignment Multiscale Adversarial Domain Adaptation YOLO for Insulator Defect Detection in Generalized Foggy Scenario

Abstract

Insulators play an important role in high-voltage systems, leading to the importance of insulator defect detection. For ideal conditions, this issue has been handled well due to the advances in objective detection methods and abundant labeled data gathered by uncrewed aerial vehicles (UAVs). However, foggy scenarios pose significant challenges: samples are difficult to be obtained and labeled, and detectors trained by clear weather samples often perform poorly in foggy scenarios. This practical dilemma for learning algorithms lies in the distribution differences between the clear weather samples obtained and foggy weather samples obtainable or even currently unseen. To decrease the differences, this study proposes attention alignment multiscale adversarial domain adaptation YOLO (A2MADA-YOLO), a novel framework that integrates domain adaptation (DA) and domain generalization (DG). It incorporates an implicit adversarial DA strategy and optimizes at multiscale. Meanwhile, an explicit attention alignment method is proposed to enhance the alignment of details. To validate its superiority, multiple experiments were conducted on state-of-the-art (SOTA) YOLO versions. The results show that the A2MADA-YOLO model significantly improved insulator defect detection performance under target domain foggy scenarios, resulting in increases in mAP0.5 for YOLOv7, v8, and v9 by 5.8%, 8.4%, and 6.4%, respectively. Moreover, this performance enhancement successfully generalized to complex foggy scenarios unseen during training, with mAP0.5 improvements of 6.5%, 8.4%, and 5.1% for YOLOv7, v8, and v9, respectively, affirming the model’s capability to learn domain-invariant features. The code is available at: <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/HaoxingZhou/A2MADA-YOLO</uri>.

Topics & Concepts