To address the challenges of low detection accuracy in aerial insulator images due to complex backgrounds, varied target scales, and small target sizes, a method for insulator defect detection called IRe-Net is proposed by improving RetinaNet. IRe-Net optimizes the ResNet network through an adaptive convolution module and enhances small target defect capture capabilities using deformable convolutions. A feature enhancement network, CeBiFPN, is designed by improving BiFPN with a coordinate attention mechanism to boost multi-scale target detection performance. Additionally, a Focal-CIoU loss function is introduced to optimize the loss function and improve model convergence, enhancing defect detection ability in complex scenes. Experimental results show that the proposed IRe-Net achieves an average accuracy of 91.46% in insulator defect detection, effectively improving detection accuracy and robustness in aerial insulator defect detection in complex scenarios, meeting the safety monitoring needs of smart grids.