Establishment and processing of Groundbased cloud image database for CNN
Wang Min1,2,Zhou Shudao1,2,Liu Zhanhua1,Ren Shangshu3
(1.College of Meteorology and Oceanography,National University of Defense Technology,Nanjing 211101,China; 2.Collaborative Innovation Center on Forecast and Evaluation of Meteorological Disasters, Nanjing University of Information Science & Technology,Nanjing 210044,China; 3.Unit 95171 of PLA,Guangzhou 510000, China)
Abstract: Convolutional neural network (CNN) has an extraordinary ability to learn features from samples,and training requires a large number of image samples with labels.Therefore,it is the first and very important step to establish cloud image sample bank when using convolutional neural network to study the groundbased cloud image.Firstly,three cloud image sample libraries are acquired by means of digital camera direct shooting,downloading from the Internet,acquiring from publicly released cloud image books,and shooting by allsky camera.Then,the resolution,noise and number of images in the three sample libraries are analyzed.Then,bilinear interpolation and data enhancement are used to normalize the sample database.Finally,CNN,LBP,Heinle feature and Textonbased method are used to verify the cloud recognition of the enhanced data set.The experimental results show that the improved data can effectively solve the problems of convolution neural network for small sample data recognition such as low rate and incomplete network operation, and lays a foundation for the application of convolutional neural network in the recognition of groundbased cloud image.
Key words : convolutional neural network;supervised learning;sample bank;normalization