<p>Risk assessment from fixed surveillance cameras is essential for the safe operation of power transmission lines. Compared with standard image classification, this task is challenging because risk is context-dependent and often determined by safety-clearance relations between small hazards and global corridor structures, while decisive visual cues often occupy only a tiny region in high-resolution images. In this paper, we propose Smart-Line, a Semantic-guided Multi-scale Adaptive enhancement framework for Risk assessment of Transmission Lines. Smart-Line harmonizes multimodal semantic reasoning with discriminative visual representations by distilling a compact prompt-conditioned semantic prior from a frozen multimodal large language model (MLLM). This prior functions as a lightweight controller to adaptively modulate and fuse multi-scale visual features through prompt-guided dynamic injection and semantic-gated feature pyramid fusion. To capture fine-grained details, a global–local view pair is constructed via prompt-based region proposals, facilitating fine-grained, risk-aware perception without the need for bounding-box supervision. To support evaluation under realistic monitoring conditions, we introduce a real-world transmission-line surveillance dataset comprising 4,606 high-resolution images with expert-verified labels. Extensive experiments demonstrate that Smart-Line consistently achieves state-of-the-art performance, yielding 0.9131 accuracy and 0.9701 AUC, surpassing the strongest baseline by 10.5 percentage points in accuracy and 7.2 percentage points in AUC. The code will be released to facilitate future research.</p>