A Deep Learning-based Method For Utility Pole Image Segmentation

Pole segmentation technology is crucial in the early detection of utility pole inclination. Due to the unclear boundaries between outdoor utility poles and their surrounding environments, existing models do not perform ideally in segmenting the boundaries of utility poles. This paper proposes an end-to-end model specifically designed for utility pole image segmentation. We introduce the Sequential Boundary Extraction Module (SBEM), which employs continuous dilated convolutions with increasing depths to obtain a larger receptive field, effectively extracting multi-scale local and global features of utility pole images, ensuring enhanced precision and completeness of edge information. Subsequently, we design the Utility Pole Positioning Module (UPPM), utilizing depthwise separable asymmetric convolutions and the SimAM attention mechanism. By leveraging the local self-similarity of deep feature maps, this module dynamically adjusts the weight of each pixel, thereby improving model performance. Finally, the model proposed in this paper is applicable to utility poles of various sizes and in different environments, achieving accurate boundary extraction, and holds potential for generalization to other types of tower-like structure segmentation tasks.