Research On Weather Image Classification Method Based On Deep Learning

All human activities in daily life are affected by weather phenomena.Intelligent applications such as autonomous driving,intelligent transportation,and intelligent security are inseparable from real-time and accurate weather information in order to maintain a stable and efficient working state.Weather image classification refers to obtaining the weather information of the location by identifying a single image.For the conventional image classification problem,there are many methods that can be well solved.However,there are still some difficulties and challenges in the classification of weather images.For example,the number and types of weather image datasets available for learning and use in academia and industry are too few to meet research needs;the training of weather image classification models requires a large number of labeled samples for support,and training with advanced computing equipment is very difficult.It takes a long time to get a model that performs well;researchers tend to focus too much on basic weather-type image classification,while ignoring the sub-classification of weather images that may be more important to people.Based on the above problems and challenges,first,this paper constructs a five-category weather image dataset that contains more categories and is closer to the real situation.The dataset includes five common categories of sunny,cloudy,rainy,snowy and foggy.Weather types,1000 annotated images per class.At the same time,based on the five types of weather image datasets,this paper also constructs a severe weather image dataset for fine-grained classification of severe weather images.Secondly,in view of the low accuracy of the existing weather image classification methods and the slow model training speed,the transfer learning method is introduced on the basis of the deep convolutional neural network,which can greatly shorten the model training time and obtain better results Classification effect.Finally,aiming at the sub-classification of severe weather images,this paper proposes a sub-classification method of severe weather images based on optimized convolutional neural network.According to certain standards,the three types of disaster weather such as rainy,snowy,and foggy weather are re-divided into six sub-categories:heavy rain,light rain,heavy snow,light snow,and heavy fog and light fog,and data enhancement is performed on them in 5 ways.,so that the number of weather categories is consistent.Then use the Swish activation function to replace the Re LU activation function,use the small convolution kernel stack to replace the large convolution kernel,and add the CBAM attention mechanism to optimize the convolutional neural network to realize the subdivision of the disaster weather.