Intelligent Ice Detection and Removal for Overhead Line Insulators: A Multi-Modal Vision Framework Integrating Adaptive Feature Fusion and Autonomous Thermal-Laser De-Icing

This paper proposes an insulator icing monitoring and de-icing method based on image segmentation. The framework integrates three key components: 1) A Linear Inverted Residual Network (LIRNet) backbone for multi-scale feature extraction, 2) A Multi-scale Feature Fusion (MFF) module to balance cross-scale feature information, and 3) An Adaptive Spatial Attention (ASA) mechanism for regional emphasis. The icing degree is quantified by calculating the ratio between insulator pixels in monitored images and reference uniced insulator images. Extensive experiments have validated that this quantification method achieves a high correlation with manual measurement results, demonstrating an average accuracy of ${9 2 \%}$ and a reliability of ${8 8 \%}$ across diverse environmental conditions, including varying illumination, weather scenarios, and viewing angles. However, challenges such as distinguishing between similar color tones of ice and insulator materials under complex lighting conditions, as well as dealing with occlusions from environmental objects, still exist and require further improvements in the feature extraction and fusion processes. When severe icing is detected, automated de-icing procedures are initiated. This approach effectively improves detection efficiency while reducing manual inspection requirements. To handle noise and outliers in the images, the method employs a preprocessing pipeline. It uses median filtering to remove impulse noise and Gaussian filtering to reduce high-frequency noise. An adaptive noise reduction technique adjusts filtering intensity based on local image statistics, ensuring a balance between noise removal and feature preservation. A statistical approach identifies and corrects outliers by analyzing pixel values within a neighborhood, maintaining the integrity of the image data. These techniques enhance the robustness of the segmentation process, leading to more accurate and stable quantification of icing degree.