Transmission line icing poses a significant threat to power grid stability by degrading both the mechanical and electrical performance of lines, potentially resulting in power outages. To address the challenges of transmission line icing classification—such as weak textures, background interference, and blurred category boundaries—this paper proposes EMA-EfficientNetV2, an attention mechanism-based convolutional neural network. Built upon the EfficientNetV2 architecture, the model integrates an Efficient Multiscale Attention mechanism to enhance feature extraction and improve classification accuracy under complex conditions. A custom dataset of iced transmission line images is utilized, with data augmentation techniques employed to mitigate overfitting. Experimental results indicate that EMA-EfficientNetV2 achieves a 2.4% increase in Top-1 accuracy and a 2.0% improvement in F1-score over the baseline model. Furthermore, the proposed approach demonstrates superior performance compared to several representative benchmark models, while preserving a lightweight architecture with a computational cost of only 2.9 GFLOPs.