Research On Lushan Mountain Landscape Image Recognition System Based On Improved YOLOv4

With the release of the tourism development plan of the 14 th five year plan,smart tourism has become an inevitable trend of tourism development.Through developing digital experience products,immersive experience services,intelligent tour and other new tourism features,tourist attractions will gradually become digital and intelligent.This will provide a new travel experience for the public.In natural scenic spots,some natural landscapes which are difficult for tourists to see rarely occur.Sometimes,tourists miss the opportunity to see the beautiful scenery when they occur because they are not familiar with the scenic spots during the tour.Therefore,it has become an important way of scenic area tourism service to generate intelligent navigation system by detecting the real-time target of landscape and conveying the viewing position to tourists.Based on the above considerations,this paper constructs the landscape recognition data set for Lushan scenic spot and proposes a lightweight neural network to meet the real-time detection speed and ensure a certain detection accuracy.Meanwhile this paper realizes the deployment of the landscape recognition system.The main work is as follows:(1)The landscape data set based on Lushan scenic spot is constructed,which includes eight kinds of landscapes such as sunrise and sunset,rime,snow,cloud sea,red leaves,mirage,Buddha light and waterfall cloud.In order to meet the needs of landscape recognition under the actual scene of Lushan Mountain,2167 landscape images under various scenes were collected by means of network acquisition and on-site image shooting.Meanwhile,the location and category of the target in each image were manually marked.Finally,aiming at the imbalance of the number of samples in the landscape dataset,the data set is expanded by the method of data enhancement.(2)A lightweight neural network algorithm is proposed to meet the detection speed and accuracy at the same time.Firstly,it uses Mobile Net V3 to replace the CSPDarknet53 of YOLOv4,so as to reduce the amount of parameters and maintain a certain accuracy.Then,while maintaining the panet structure and spp structure,the ordinary convolution is replaced by deep separable convolution to speed up the detection speed of the network.Finally,by comparing with the lightweight neural network which is widely used now,it shows that the detection speed of the algorithm is enough to meet the needs of real-time detection while ensuring relative detection accuracy.(3)A landscape recognition system platform is built and the corresponding online recognition system software is designed and developed.The system realizes the real-time recognition of landscape in the actual environment and tests the corresponding pictures to verify the feasibility of the proposed lightweight neural network.