Three-channel feature fusion network for weather image recognition
Zheng Yuan, Chengyin Ye · 2025
In view of the poor recognition effect of weather images under different scenes and conditions and the difficulty in recognition caused by similar weather phenomena, this paper proposes a three-channel feature fusion network VGG19_F to recognize weather images. First, five traditional features of color histogram, contrast, saturation, LBP, and dark channel are extracted as traditional channel features; then the fully connected layer of the VGG19 network is replaced by a self-designed fully connected layer, and the extracted features are used as fully connected channel features; then an average pooling layer is added after the last convolution block, and the extracted features are used as average pooling channel features; finally, the features of the three channels are fused, and the classifier is trained with the fusion features to classify weather images into sunny, cloudy, rainy, snowy, foggy, and lightning. Experimental results show that the three-channel feature fusion network can achieve a recognition accuracy of 95.65%. This model can combine the advantages of traditional methods and deep learning methods to improve the accuracy of existing weather image recognition methods, while improving the recognition rate under weather phenomena with similar features.