Fine-grained Classification Of Weather Types Based On Outdoor Scene Images

With the growing development of artificial intelligence,many outdoor visual systems are gradually improving and developing.Therefore,real-time weather conditions also affect outdoor monitoring,weather prediction,disaster warning,automatic driving,scene understanding and other fields based on image data.To make the computer more anthropomorphic and intelligent in weather classification through images has become one of the research subjects in the field of computer vision.Based on this,this paper uses the deep learning network to build the relevant classification and prediction model from the two directions of basic weather type classification and severe weather fine-grained classification,so that the computer can complete the basic weather type and severe weather type classification through data-driven automatic learning.For the classification of basic weather type,this paper designs a deep network framework with dual attention mechanism based on image semantic segmentation.To be specific,firstly,the weather image is semantic segmented to obtain the sky and ground regions of an image.Then,the features of these two regions are learned,and the learning process is completed by the residual network.Secondly,in order to extract the weather type features with discriminative information,the channel attention and spatial attention learning are performed in the deep network,and the obtained channel features and spatial features are cascaded.Different attention is paid to the feature map from the two dimensions of channel and space respectively,focusing on the extraction of useful information,so as to obtain the strong and effective discriminant ability of the feature map.Finally,the proposed algorithm is compared with other basic weather image classification algorithms,and the results show that the proposed algorithm can achieve more than 99%accuracy on four and six basic weather data sets respectively.In this paper,an improved residual network model based on data enhancement and channel attention is proposed for the fine-grained classification of severe weather images.Specifically,five kinds of data augmentation operations are performed on the original image,including rotation,flip,noise,shift and clipping,to expand the image data of the category with fewer samples and make the data more balanced.Moreover,Resnet20,a residual network design with a simplified number of layers,is considered to better complete feature learning.Then,channel attention mechanism is used to further make discriminant characterization of features.Finally,creating the task of weather datasets,a fine-grained image dataset containing heavy and light levels of rain,snow,and fog.Experiments based on these datasets show that our proposed fine-grained classification network model for severe weather is effective on self-built data sets,and can achieve more than 80% accuracy.