基于残差Swin Transformer的天气图像识别技术研究

Research on Weather Image Recognition based on Residual Swin Transformer

人类活动经常受到天气条件的影响,基于图像的自动天气识别在实际应用中具有重要意义。然而现有方法均使用卷积神经网络,未能有效地利用图像的全局信息和像素点之间长距离的依赖关系,且识别的天气类型较少,识别精度较低。为解决这些问题,尝试将视觉Transformer应用到天气识别领域,同时提出一种基于残差Swin Transformer的模型,并使用先进的优化器Ranger来提高天气识别的正确率。该模型在包含11种天气现象的公开数据集WEAPD上进行验证,实验结果表明,其整体性能优于其他先进的识别网络,识别正确率达到93.6%,可为天气图像识别和天气预报研究提供参考。

Human activities are often affected by weather conditions,and automatic weather recognition-based image is of great importance in practical applications.However,existing methods all use convolutional neural networks,which fail to effectively utilize the global information of images and the long-distance dependency between pixel points,and recognize fewer weather types with low recognition accuracy.To solve these problems,we try to apply the visual Transformer to the field of weather recognition,and also propose a model based on the residual Swin Transformer and use the advanced optimizer Ranger to improve the weather recognition rate.The model is validated on WEAPD,a publicly available dataset containing 11 weather phenomena,and the results show that its overall performance is better than other advanced recognition networks,with a 93.6%correct recognition rate.It can benefit the research of weather image recognition and weather forecasting.