Scene text detection has attained substantial advancements under normal weather conditions. However, its accuracy significantly degrades in adverse weather, such as snowy conditions, making it unsuitable for practical applications. To alleviate the influence of snowy weather on text detection, we propose a dual-branch architecture, termed SnowTextNet. We first design a novel desnowing network that establishes a positive correlation between image desnowing and text detection by incorporating guidance from text detection masks and constraints by detection loss. Moreover, to reduce negative impact of poor desnowing results on the detection, we adopt dual-branch architecture and design a feature fusion module to combine the key features from the original snowy image and the relatively clear features from the desnowed image. Furthermore, we propose a two-stage training strategy. Initially, we train a detection-friendly desnowing network. Subsequently, the desnowing and text detection modules undergo joint optimization using the dual-branch architecture. Following this training strategy, SnowTextNet attains the best detection performance on snowy scene datasets.