Improved MViTv2-T model for insulator defect detection

Abstract

<p>Insulators play a crucial role in transmission lines. Insulators exposed to natural environments are prone to various malfunctions. These faults will seriously affect the safety and stability of the power grid system operation, so intelligent detection of insulator defects has become increasingly important. This paper presents an insulator defect detection model based on the improved MViTv2-T (Multiscale Vision Transformers Version 2 Tiny). The new model utilizes the sore penalty mechanism (SPM) cluster non-maximum suppression (NMS) algorithm instead of the batched non-maximum suppression (NMS) algorithm from the original model. Additionally, it introduces the stage query recollection method, which integrates high-level and low-level module queries within each stage, along with various experimentation on integration functions between the two. The experimental results indicate that the improved MViTv2-T model attains an mAP (mean average precision)@0.5:0.95 of 76.1$ \% $, mAP@0.5 of 96.1$ \% $, and mAR@0.5 of 97.2$ \% $ in insulator defect detection. Compared to the original model, there is a 1.8$ \% $ increase in mAP@0.5:0.95 and a 17$ \% $ decrease in the detection error rate at an Intersection over Union (IoU) threshold of 0.5. Furthermore, when compared to standard two-stage detection models and YOLO series models, the improved MViTv2-T model also exhibits distinct performance advantages.</p>