Power transmission lines are a critical component of the entire power system, and ice accretion incidents caused by various types of power systems can result in immeasurable harm. Currently, network models used for ice detection on power transmission lines require a substantial amount of sample data to support their training, and their drawback is that detection accuracy is significantly affected by the inaccurate annotation among training dataset. Therefore, we propose a transformer-based detection model, structured into two stages to collectively address the impact of inaccurate datasets on model training. In the first stage, a spatial similarity enhancement (SSE) module is designed to leverage spatial information to enhance the construction of the detection framework, thereby improving the accuracy of the detector. In the second stage, a target similarity enhancement (TSE) module is introduced to enhance object-related features, reducing the impact of inaccurate data on model training, thereby expanding global correlation. Additionally, by incorporating a multi-head adaptive attention window (MAAW), spatial information is combined with category information to achieve information interaction. Simultaneously, a quasi-wavelet structure, compatible with deep learning, is employed to highlight subtle features at different scales. Experimental results indicate that the proposed model in this paper outperforms existing mainstream detection models, demonstrating superior performance and stability.