Weather Image Classification Based On Data Augment

The problem of weather classification plays an important role in the fields of automatic driving,outdoor detection and agricultural field.Weather image classification is a special image classification problem.Different from animal or plant image classification,the main body of weather image may contain various objects.The classification model needs to judge the weather by clues in image.The complexity of weather classification problem does not only come from the diversity of clues in images.In real world,there are often many kinds of weather coexisting at the same place and at the same time.Multi-label classification increases the difficulty of weather classification.This thesis discusses the single label weather classification task(one picture corresponds to one weather)and the multi-label weather classification task(one picture corresponds to multiple weather).It mainly studies the weather image classification methods based on data augment.The main content is as follows:1.For the single label weather classification task,this thesis proposes an improved Res Net(Residual Network)image classification network based on attention mechanism and global color extraction module and RBFCGAN(improved condictional generative adversial net-work based on RBF kernel)feature enhancement network.In this thesis,the image clas-sification network is logically divided into feature extractor and feature classifier.The image data set is processed into feature set by the feature extractor,the feature set is enhanced by RBFCGAN,and the enhanced feature set is used to fine-tune the feature classifier,so as to improve the classification ability of the classification model.2.For the multi-label weather classification task,this thesis proposes an adaptive knowledge distillation method to improve the generalization ability of the classification model CAM-Res Net(Res Net using CAM module).It uses the classification model to generate CAM(class activation feature map)to mark the image areas which strongly associated with the category,and proposes a local Mix Up method to fuse and generate multi-label weather image samples to make up for the imbalance of multi weather dataset,so as to improve the performance of the classification model on multi-label data task.3.In view of all kinds of cumbersome work in model training and data management,this thesis designs and implements the weather classification model and data management and training system.The system can manage the model and data more efficiently,pro-vide real-time data visualization and interactive behavior during training.The system improves the work efficiency of researchers,provides relevant staffs with convenient training and deployment functions of weather classification model.To sum up,this thesis studies the classification task of weather image,and extends from the relatively simple single label weather classification problem to the complex multi-label weather classification problem.The experimental results show that the data augment methods and image classification models proposed in this paper have good results in the weather image classifica-tion task.The weather classification model and data management and training system imple-mented in the thesis has a friendly interface and can provide relevant staff with convenient training and deployment functions of weather image classification model.