Current scene text recognition methods perform well on clear images but face challenges under adverse weather conditions such as snow. Although image desnowing can serve as a preprocessing technique, existing desnowing models primarily target natural scenes and fail to account for text-specific characteristics, resulting in limited effectiveness for text image restoration. To address this, we propose Text-Focused Snow Removal Network (TFRNet), which incorporates a dedicated snow removal module and connects the restoration and recognition modules through an end-to-end framework to enhance text recognition performance in snowy environments. The restoration module achieves text structure perception by fusing global and local features. Additionally, we design a Multi-Scale Sequential Residual Attention Block (MSRAB), which enables us to build a correlation in the fore-and-aft characters. Experimental results demonstrate that TFRNet significantly improves text recognition performance in snowy conditions and achieves superior performance on both synthetic and real-world datasets.