A Street View Image Change Detection Method Based On Fully Connected Neural Network

Street scene image change detection can detect change information from different perspectives,and has a unique facade detection method,which can quickly detect changes in urban structure and street scene,therefore,it has important research significance in the fields of smart city construction,urban management,disaster damage assessment and disaster emergency management.However,street scene images have the characteristics of complex scenes,traditional change detection methods and machine learning methods have the problems of low overall accuracy and high false detection rate.Fully connected neural network can process images according to the texture,scene and other high-level semantic features,it has been widely used in image processing and computer vision.The use of a fully connected neural network can solve the problems of "same thing with different spectrum" and "different thing with the same spectrum " in street scene images to a certain extent.Therefore,this paper studies the street view image change detection method based on fully connected neural network.The main work and research contents of this paper are as follows:(1)The semantic segmentation data set of street scene image is established.The street scene image change detection method based on fully connected neural network used in this paper is based on two phases of street scene image semantic segmentation,but currently there are fewer training data sets available for street scene semantic segmentation,and it is difficult to meet different street scene semantic segmentation.For this reason,this article creates a street view image semantic segmentation dataset based on the labelme software,and uses it to augment the Camvid dataset;(2)An improved DeeplabV3+ network street scene image change detection method is proposed.Aiming at the problems of insufficient generalization ability and low detection accuracy of DeeplabV3+ network migration learning model,an improved DeeplabV3+ network street scene image change detection method is proposed.This method optimizes the training data and improves the network structure,then selects the optimal network model for street view scene change detection.The experimental results show that the overall accuracy of the two groups of data obtained by this method is 82% and 63% respectively,the kappa coefficient is 49.39% and 26.36%respectively,and the F-measure coefficient is 89.4% and 79.3% respectively,the results are better than other comparison methods;(3)A street scene image change detection method based on DeeplabV3+ network combined with graph cuts is proposed.There is a problem of "same thing with different spectrum" in the sky of street scene images,and its change detection result seriously affects the change detection result of DeeplabV3+ network migration learning method.Based on this,a street scene image change detection method based on DeeplabV3+network combined with graph cuts is proposed in this paper.graph cuts is used to eliminate the influence of sky and vegetation in this method,thereby reducing the false detection rate and improving the accuracy of change detection.The experimental results show that the overall accuracy of the two groups of data obtained by this method is 77% and 67% respectively,the kappa coefficient is 43.14% and 25.06%respectively,and the F-measure coefficient is 86.1% and 81.5% respectively,which is better than other comparison methods.