Research on the application technology of transmission line icing segmentation detection model based on Mamba’s visual neural network

In mountainous areas with high humidity and cold temperatures, the icing of transmission lines is prone to cause cascading risks such as conductor overload, sag increase, and galloping instability, which are the main external causes of power grid failures in winter. Traditional manual inspection and threshold image algorithms are difficult to achieve efficient, real-time, and low-cost icing monitoring under complex terrain, low illumination, and snow and fog interference. To this end, this paper proposes a lightweight icing instance segmentation framework: embedding bidirectional VMamba state space blocks into the YOLOv12 backbone network to balance global dependency modeling and linear computational complexity. Experiments show that the model achieves an mAP of 86.3% with only 12.5 M parameters and an inference speed of 66 FPS, which is 3.7 percentage points higher than the lightweight network YOLOv12s of the same level and significantly better than traditional two-stage methods. The model achieves an mAP of up to 91.2% in low-light scenarios and maintains ≥85% mAP under geometric deformations such as translation, rotation, and flipping, as well as optical degradation such as low light and exposure changes, demonstrating excellent environmental adaptability. The main advantages lie in high accuracy, high real-time performance, low resource consumption, and strong robustness; its limitations include the need for further improvement in performance under extreme optical interference such as Gaussian blur and channel distortion. Overall, this work breaks through the contradictory triangle of 'high resolution-real-time performance-low power consumption', providing a deployable and scalable efficient solution for online icing monitoring of transmission lines. The theoretical basis for this method to break through the "high resolution-real-time performance-low power consumption" triangle of contradictions lies in: (1) VMamba’s SS2D module reduces the computational complexity from O(N2) to O(4N), with only 32 FLOPs at an input of 1280 × 720; (2) Cross-direction scanning enables continuous memory access, reducing the number of DRAM accesses by 63.2%; (3) The dynamic resolution mechanism enables sub-linear growth of power consumption with input size, with a power consumption of 15.1 W at an input of 768 × 768, a reduction of 22.7% compared to the fixed resolution strategy. Highlights