Classification of Multi-Class Weather Image Data

Abstract

In order to improve the accuracy of multi-class weather image recognition, a weather image classification algorithm based on ViT was proposed. The multi-layer transformer method is used for feature extraction and encoder. The self attention method is introduced to calculate the similarity between each position in the segmented image dataset. Then, all positions are weighted according to the score to generate a new feature representation. Finally, the weather image data is classified by MLP classifier. Experiments show that the deep learning algorithm based on ViT can effectively improve the prediction accuracy of the model, and has a good effect on improving the recognition of multi-class weather images. The algorithm can achieve good accuracy on 14 different weather data, with a recognition accuracy of up to 92.83%.