Research On UAV Aerial Image Difference Detection Algorithm Based On Deep Belief Network

The difference detection technology of UAV aerial images is of great practical significance for acquiring the dynamic changes of the city in time and implementing the reasonable planning and management.UAV image has the advantages of high resolution and fast acquisition,which can provide timely and effective data source for urban development research and realize continuous and dynamic urban difference detection.With the rapid development of deep learning technology,deep neural network can automatically learn the advantages of deep features of complex data,which provides a new research idea for difference detection.Therefore,it has a certain engineering application value to carry out the research of UAV aerial image difference detection algorithm based on deep learning in this application background.In this paper,aiming at the disadvantages of high-resolution UAV aerial image,such as large amount of data,low matching efficiency and long time-consuming,a UAV aerial image registration algorithm based on secondary matching is proposed.Firstly,the image dimension is reduced by down sampling,and coarse matching and fine matching are combined by orb algorithm to achieve the registration of high-resolution UAV aerial image.Experimental results show that the algorithm can significantly reduce the registration time and improve the registration efficiency.Aiming at the problem that the training of neural network needs a lot of label data,but the manual annotation of high-resolution UAV aerial image is heavy,inefficient and the label data is insufficient,a difference detection algorithm of UAV aerial image based on weak supervised deep belief network(DBN)is proposed.This algorithm takes the pre classification result map after median filtering as the pseudo label of network training To build and train the difference detection network model based on DBN to realize the final detection.The experimental results show that the algorithm in this paper has a certain feasibility,and in the process of implementation,it does not need to mark the real difference information artificially.It realizes the automation and intelligence of the difference detection,and effectively improves the detection accuracy.