Aiming at the problems of inaccurate positioning, low recognition efficiency, and difficult segmentation of insulator image in complex background, an insulator infrared image segmentation algorithm based on the dynamic mask and box annotation is proposed. In the conditional convolution instance segmentation network framework, the single-shot and high-performance box-supervised methods are introduced, and the mask-level loss is used to replace the pixel-level loss. By marking the insulators in the infrared image with a rectangular frame, the algorithm can realize the overall segmentation of the insulator string. The experimental results show that the algorithm uses the ResNet-101-FPN backbone and 3x training plan to train a large number of infrared images, which can achieve accurate identification and high-precision segmentation of insulators, and demonstrate excellent segmentation performance.