The REGDETR (reg detection transformer) model is proposed to address the diversity and complexity in insulator defect detection.This model is based on the end-to-end RTDETR (realtime detection transformer) architecture,and three new modules are introduced to enhance detection performance.First,the Rep-Faster-Basic-Block module is proposed,effectively capturing multi-scale features using reparameterized T-shaped convolutions.Second,the ELA-Basic-Block (efficient local attention basic block) module is employed to integrate high-dimensional feature representation.Finally,the GS-HWD-BLRA (gsconv Haar wavelet downsampling with bi-level routing attention) module is introduced to replace the stride convolution in the feature fusion network.This module retains image information using the Haar wavelet transform and filters redundant information through a bi-level routing attention.Experimental results show that REGDETR achieves a 4.5 % improvement in mean average precision,reaching 95.7 %,while maintaining the same parameter count and significantly reducing computational load,surpassing existing similar models and enhancing the accuracy of insulator defect detection.