Research On Garbage Scattered Areas Detection In Low Altitude Remote Sensing Images Based On End-to-End Deep Learning

A large number of different ideas,models and methods have been proposed for the research of change detection using remote sensing images.However,how to use the low-altitude remote sensing images of small UAVs for fast,accurate,and end-toend change detection research is still a difficult problem.The bi-temporal low-altitude remote sensing images taken by small UAVs are often affected by positioning error and wind interference.The difference in view point will directly affect the result of change detection.In addition,it is also a challenging task to use computer vision technology to replace human decision-making,and to detect garbage scattered areas on the ground through low-altitude remote sensing images of nature reserves.To address these problems,a change detection network for garbage scattered areas in nature reserves is proposed by using bi-temporal low-altitude remote sensing images of small UAVs,which contains the following contributions.(1)End-to-end change detection network: an end-to-end change detection network without additional preregistration network is concatenated by an optical flow registration network and a change detection network;(2)Siamese network down-sampling: the pre-trained Res Net18(Residual Network 18)network is used as the backbone,and the Siamese network is constructed by the shared weight method to complete the down-sampling;(3)Optical flow pyramid: global correlation in the deep layer of the registration subnetwork,and local correlation in the shallow layer of the network are used to build optical flow pyramid,from coarse to fine layer-by-layer optical flow field for feature map registration;(4)Up-sampling nested connection: skip connection is used to connect features at the same feature level The feature fusion of the graph is performed,and new up-sampling paths are constructed to fuse the feature maps of different feature levels,so that the fused feature map contains both shallow position information and deep semantic information;(5)Channel group attention mechanism: based on the channel attention and spatial attention in the convolutional block attention mechanism CBAM,the feature maps obtained by different feature map paths are first subjected to intra-group and inter-group channel attention,and then global spatial attention method to optimize the representation of semantic information and location information in feature maps.The data of the study mainly consists of 2,000 images of different nature reserves captured by small UAVs,and a multi-label dataset containing 48,000 images was generated through data augmentation.This study evaluates some change detection state-of-the-art methods,and the datasets used are all pre-registered with GLU-Net.The experimental results show that our method performs better in accuracy and performance by comparing with eight change detection algorithms(FC-Siam-Conc,FC-Siam-Diff,UNet++＿MSOF,IFN,DASNet,BIT,SNUNet-CD,RDP-Net),and has better real-time performance and integrity.