Defective insulator detection is one of the critical tasks in the operation and maintenance of smart grids. To address the challenges of multi-target and multi-scale detection in aerial insulator images, a defective insulator detection method is proposed based on an improved CenterNet architecture. The method adopts an anchor-free detector as the foundational framework and innovatively integrates three key technologies. Firstly, an expanded feature enhancement module is designed, which effectively enlarges the feature receptive field through dilated convolutions, significantly improving the model's ability to capture multi-scale target features. Secondly, a convolutional block attention mechanism is embedded into the network to dynamically optimize the weight distribution of feature channels, enhancing both detection accuracy and computational efficiency. Finally, a multi-scale feature pyramid structure is employed to achieve the fusion and complementarity of multi-level features. Experimental validation demonstrates that this method excels in defective insulator detection under complex scenarios, achieving an average precision of 95.17 %, with all metrics significantly outperforming existing mainstream algorithms, fully proving its advantages in practical applications for power line inspection.