In response to the challenges of detecting small targets, large scale differences, and complex backgrounds in aerial images of insulator defects in power transmission lines, we propose an improved insulator defect detection model, PDS-YOLO. First, the C2f module was improved by replacing the traditional bottleneck module with Poly Kernel Inception Network (PKI) Block, which effectively reduces the loss of detail information in complex backgrounds, thereby improving the accuracy of small object detection. Next, we design the Dynamic Fusion upsampling (DFsample) module, which employs a dynamic sampling strategy to optimize the upsampling process, improving the ability to restore details. Finally, we replace the CIoU loss function with the SIoU loss function, which improves the precision of small target detection. Experimental results show that PDS-YOLO achieves a mAP50 of 92.5% and a mAP50-95 of 66.7%, which are improvements of 2.4% and 2.2%, respectively, compared to the baseline YOLOv8 model. These results demonstrate the superiority of PDS-YOLO in insulator defect detection, particularly in small target detection and multi-scale target handling, where it shows significant advantages.