Research And Application Of Outdoor Weather Image Classification Method Based On CNN And Transfer Learning

Timely and accurate weather information helps intelligent systems such as autonomous driving,intelligent monitoring,and intelligent transportation to make optimal decisions.This article will study how to obtain the weather information of the current environment from a single image.This research field is called weather image classification.At present,the general image classification problem has a better solution,but the weather image classification task still faces many challenges.For example,the size of the weather image data set is small,and the model is difficult to train and optimize;there are a large number of the same targets and features in images representing different weather conditions;the problem of fine classification of weather images is ignored.Based on the above problems and challenges,first of all,this paper builds four types of weather image data sets(FWID)for outdoor transmission line scenes.This data set contains four types of weather: foggy,rainy,snowy and sunny,with a total of5395 marked images Based on the four types of weather image data sets,this paper also builds a block weather image data set(BWID)and a subdivision weather image data set(SWID).Secondly,aiming at problem of small scale of weather image data sets and difficulty in training models,this paper designs a weather image classification method based on deep transfer learning;this method applies transfer learning technology to the deep convolutional neural network model,which can greatly improve the model Training optimizes speed.Then,in view of the problem that there are a large number of the same targets and features in images representing different weather conditions,this paper proposes and implements a weather image classification method based on block images and voting strategies;this method cuts a complete image into several Small blocks,and then use the convolutional neural network to extract the weather-specific features(snow,raindrops,fog)in the image,and finally use the threshold-based voting strategy to get the weather category;this method is achieved on the weather image dataset constructed in this paper 98.740.3% classification accuracy.Finally,for the fine-grained classification of weather images,this paper designsand implements a fine classification method of weather images based on multi-task learning.Taking snow days as an example,this paper divides them into two sub-categories of first-level snow days and second-level snow days according to certain standards,and then combines the parameter hard-sharing multi-task learning model to achieve clear,foggy,rainy,and Distinguish between first grade snow days and second grade snow days.