Research On Weather Recognition Methods Based On Feature Selection

Recently,intelligent equipment not only influences us in our daily lives through replacing the human to complete a variety of difficult and high-risk tasks,but also brings convenience.Complex and varied weather conditions,such as rainy,haze and blizzard and so on,strongly influence the general functionality of vision systems.Weather recognition is an important and challenging task in the fields of computer vision and pattern recognition,which automatically analyzes the weather conditions in the scene and provides effective protection for the event judgment.At the same time,the accuracy of outdoor video surveillance is improved and the cost of traditional weather sensors is reduced.Weather recognition technologies are applied widely in many areas,such as intelligent video surveillance system,machine vision navigation and vehicle driving support system.Hence,research of weather recognition has remarkable value and application prospects.The traditional weather recognition uses the statistical analysis to recognize the existing weather data,which has three problems for weather identification and prediction: a longer observation time,large amounts of samples and non-real time.At present,weather recognition based on images is being studied.This method extracts different image features(color,texture,etc.)to describe weather conditions.Generally,weather images can be described by multi-feature fusion for obtaining better recognition results,but higher dimension after multi-feature fusion will reduce the efficiency of algorithm.In addition,using a single classifier can't distinguish similar weather images between cloudy and overcast well.Regarding the issues mentioned above,this paper presents two weather recognition methods named MGMKL(Multi-feature Generalized Multiple Kernel Learning)and MCSVM(Multi-feature Compositional Support Vector Machine)respectively.Feature extraction is one of the key steps in our method that extracts seven groups of frequently-used features,including SIFT,HSV color,LBP,Contrast and Gradient,and then using Bag-of-Feature model(BOF)to form feature vectors for to describe weather images effectively.In MGMKL,the feature weights are continually optimized for each feature vector after seven features integration,and then we selects good component corresponding to the larger weight to express content of weather images.MCSVM integrates multiple weak classifiers(SVM)for determining the importance of each set of features,and selecting the best combination to describe the weather images.The proposed two feature selection methods can effectively reduce the feature dimension,and achieve good performances for weather recognition.The proposed methods is tested on three weather recognition databases,and achieves higher recognition rate.The panorama scene and traffic scene dataset are publicly standard datasets.The fixed angle data contains 1000 images which is collected based on the University of Arizona's laboratory,including three kinds of weather conditions(sunny,cloudy and overcast).Compared with two existing weather recognition methods,experimental results show that our methods can efficiently recognize weather.