The precise localization of insulators in visible images is crucial for ensuring the integrity of power systems. In UVAs inspections, challenges such as arbitrary angles, dense distributions, and complex backgrounds often hinder accurate detection. However, existing methods primarily employ axial horizontal detection boxes to detect insulators, which overlook insulator characteristics. To further improve the accuracy, this paper optimizes YOLOv5 across four key areas: rectangular box representation, feature extraction, lightweight convolution, and detection head design. The results show that this paper uses rotating detection boxes instead of axial horizontal detection boxes, which removes a large amount of redundant background information and improves the average accuracy and operation speed. Our experiments on the insulator defect dataset achieved the average accuracy of 97.91% at an intersection ratio of 0.5. Additionally, our algorithm shows promise for extending to defect detection in various power equipment, laying a solid foundation for future automation initiatives.