Research On Construction Scene Recognition By Convolutional Neural Network Based On UAV Images

Construction in the rail transit protection zone brings hidden safety hazards to rail transit operations,resulting in frequent accidents.At present,the inspection is mainly carried out by manual periodic inspection.This inspection method has the problems of high labor cost,poor timeliness,and undetectable construction in blind vision areas.In order to solve this practical problem,a method of patrolling the rail transit protection area using drones is proposed.The key to achieving UAV inspection is the recognition of construction scenes in drone images.In view of the good performance of convolutional neural network in image recognition and classification,different convolutional neural network models are used to identify the construction scenes in UAV images,and the effects of different models are compared and analyzed.In order to meet the training needs of convolutional neural network,the images are made into labeled data set firstly.Four classic convolutional neural networks,Le Net-5,Alex Net,VGG16,and Res Net50,are trained on the data set.The comparative analysis obtains the following conclusions: the deeper the network,the higher the accuracy,and the accuracy of VGG16 and Res Net50 is significantly higher than Le Net-5 and Alex Net;but the deeper network means longer model training time and higher hardware configuration,the classification accuracy of Le Net-5 can reach more than85%,so it is a more practical choice in the case of low accuracy requirement.In the same network,setting different parameters will also result in different effects.The setting of the batch size,the adjustment of the learning rate,and the selection of the optimizer all need to be studied.The experimental results show that for the specific application of identifying the construction scene in the drone image,when the batch size is set to 64,a good balance can be achieved between the memory size and the training speed;the learning rate adjusted according to the number of training rounds is better than the method adjusted according to the loss of the verification set;the use of the Adam optimizer is the best among the three optimizers SGD,RMSprop,and Adam.Convolutional neural network training requires a large amount of data.In order to increase the number and diversity of samples,it is often necessary to use various data enhancement methods to process the collected data.Different data enhancement methods have different effects in actual use.Using affine transformation,brightness transformation,contrast transformation and Gaussian blur to expand the sample and compare the training effect of Le Net-5 on the expanded sample,it is found that the sample with increased brightness has the most obvious improvement on the training effect of the model.Rotating the image at a small angle can also achieve good results.This article describes the data processing flow of using a convolutional neural network to identify construction scenes in drone images.Experiments prove the performance differences of different networks and the impact of different parameter selections and different data enhancement methods on the training process,providing a possible choice for automated patrols in rail transit protection areas.