Instance Segmentation Algorithm Based on Improved YoloV8

In power systems, the safe and stable operation of transmission lines is of paramount importance. Insulators, as the critical component of transmission lines, play a direct role in ensuring the reliability of the power grid. As the scale of transmission lines continues to expand and the operating environment becomes increasingly complex, traditional manual inspection methods can no longer meet the requirement for high efficiency and high precision monitoring. Therefore, automated and intelligent monitoring technologies have become an important development direction for power system maintenance. To achieve a high-precision, lightweight segmentation network, improvements to the YOLOv8 network are proposed. First, a backbone network named SubNet is proposed, which has a smaller parameter count and stable accuracy. Second, the Bottleneck in the C2f module of the neck network is optimized to a star connection module (Star_block), constructing the C2f_Star structure to reduce model complexity and minimize redundant calculations. Finally, the segmentation head module is improved by introducing a new segmentation head (LSCD-Segment), further reducing the number of model parameters and enhancing detection performance. Experiments show that, compared to traditional object detection models, there is a 50% reduction in parameter count while maintaining relatively stable accuracy. This allows for better detection performance with fewer parameters, making it more suitable for deployment on mobile devices and promoting the development of drone power inspection technology.