Visual Image Weather Recognition Based On Convolutional Neural Networks

The identification and prediction of weather conditions are of great significance to the fields of traffic safety,environment and meteorology.Under the technical background that all kinds of industries transform to intelligent,an efficient method of weather recognition based on artificial intelligence technology can not only solve the problem of low accuracy of traditional weather recognition methods,but also can realize the real-time judgement to improvement the ability to cope with various of weather conditions.Convolutional Neural Network(CNN)is an important network structure in deep learning.By introducing convolutional layer,pooling layer and deeper network structure,CNN perceives higher semantic features and enhances the image classification effect.In this paper,based on the convolution neural network architecture,the corresponding weather recognition framework is studied in view of the weather conditions of visual images(sunny,foggy,rainy,snowy)which are difficult to be identified by the traditional weather recognition methods,and we also compared the method with the traditional methods using the traffic video images.The main works are as follows:1)Considering the high similarity of the weather features between sunny and snowy images,and rainy and foggy images,the recognition rate of the method using CNN alone is lower,we studied a CNN weather recognition method based on edge degeneration.This method used Mask R-CNN to extract the foreground part and the foreground edge,after extraction,the edge,foreground,background and the whole image were converted to the same scale and superposed into a three-dimensional matrix.Then CNN was used to extract and classify the features of the three-dimensional matrix.2)A weather recognition system is designed and implemented to validate the proposed method used on the image data obtained by traffic camera and the WILD database.Compared with the existing methods of weather recognition of visual images,this method can greatly improve the recognition accuracy of the four weather conditions from 78.6% to94.7%.