A Study Of Multi-class Weather Image Classification Algorithm

The weather conditions affect our daily lives in many ways,such as agriculture,aquaculture and transportation.In particular,extreme weather always brings potential risk to driving,which leads to people's life and property being put into great dangers.In recent years,governments and consumers have been deeply aware of the importance of reducing the frequency of accidents and mitigating accidents,consumers are paying more and more attention to car safety.Therefore,instantly and densely collecting weather information is a scientific topic with enormous social impact.The automatic recognition of weather plays an important role in the application of the traffic condition warning,automobile auxiliary driving,climate analysis and so on.To improve machine vision in adverse weather situations,a reliable weather conditions detection system is necessary as a ground base.Image weather recognition is a relatively new topic in computer vision.Different from object and scene recognition problems,weather recognition needs to understand complex phenomena of lighting and reflection on object and scene surface.Multi-class weather image classification remains a challenging task due to the diversity,variability and high dependence on each other of weather feature.Most of the existing methods in computer vision are based on the assumption that the outdoor weather images or videos is clear and discernible,which causes the limitation performance of the existing weather classification model on more complex adverse weather images.This paper focuses on the problems of the existing weather image database and the extraction of more effective weather features,the main contributions are summarized as follows.Firstly,we newly constructed a multi-class weather image dataset(AMWI).It contains sunny,cloudy,overcast,rainy,snowy,and hazy weather images which has more adverse weather conditions,including real-time weather index features corresponding to weather images.In order to solve the problem of the imbalance of the number of images between different weather classes in the weather image dataset,we use cycle-consistent adversarial networks to convert weather image classes using sunny images which has a large number and more diverse backgrounds to expand samples of other weather images.Because of the lack of the railway transportation weather image dataset in the driving assistance system based on the weather image classification,we also constructed railway transportation dataset RTWI from the real videos which catains sunny and cloudy images.Secondly,we proposed two weather image classification models to capture more discriminate feature for each weather condition.Firstly,we proposed a multi-class weather classification method fusing weather index features and image features.It combined the advantages of weather image features and weather index features to improve the accuracy of weather image classification.Next,we provide a multi-task learning framework which formulates the classification problem as a multi-task regression problem by considering the classification on each weather class as a task.We capture the group structure among features by a group Lasso regularization to select features and improve the accuracy of weather classification.To evaluate the performance of our proposed framework in traffic scene,we also conduct experiments on railway transportation dataset.