Addressing the issues of insufficient precision and difficulty in identifying small targets in deep learning-based aerial image power grid facility segmentation methods, this paper proposes an improved framework integrating multi-scale features and boundary awareness based on SAM. A dual-branch architecture combining ResNet-34 and ViT is constructed within the Image Encoder, balancing global structure capture and local detail extraction through channel-wise concatenation. An adaptive prompting mechanism is established leveraging the Feature Pyramid Network (FPN) and Adaptive Prompt Generator (APG), enhancing the automation and accuracy of power grid segmentation. A Boundary Awareness Module (BAM) and Channel-Spatial Hybrid Attention (CBAM) are embedded in the Mask Decoder, paired with a cross-entropy-edge hybrid loss function that strengthens the boundary segmentation accuracy of small targets by constraining both boundary precision and recall. Ablation studies and comparative experiments with different algorithms validate the effectiveness and advancement of the proposed enhanced SAM model for aerial image power grid segmentation tasks.
Guo et al. (Wed,) studied this question.