Research On Deep Learning Segmentation Method Of UAV Searching For Physical Evidence Images

Searching for on-site physical evidence plays a key role in correctly analyzing the case and finding out the facts of the case.At present,the scene environment is very complex for most cases.It’s difficult to obtain the evidence manually of the cases that occured in the suburbs,mountains,et al.So the use of Unmanned Aerial Vehicle(UAV)can greatly improve the search efficiency.However,the information of the physical evidence in the images captured by UAV is easily affected by complex environments such as grassland.And subjectively,it is difficult for people to find and judge the physical evidence accurately through their eyes.Therefore,the semantic segmentation is used in the paper to segment the physical evidence of the scene to assist the search.Based on the demand of accurate segmentation of physical evidence images and efficient segmentation in UAV search,the segmentation accuracy and speed are studied in the paper.The specific work contents are as follows:1)In terms of the segmentation accuracy of physical evidence,an improved physical evidence segment method based on Deeplabv3+ is proposed.It can solve the problems of incomplete object segmentation and poor edge recognition in the segmentation of the high-precision images.In the encoder,the dilatation rate of dilated convolution in ASPP module is reduced and embedded in the SE-block,which is called SE-ASP module.In the decoder,the output features of the encoder are fused layer by layer to improve the edge segmentation effect of the physical evidence.The Focal loss function is used to improved the problem of the sample imbalance.The experimental results show that the improved method improves m PA by 5.05% and m Io U by 2.91% compared with Deeplabv3+.The segmentation results of the physical evidence exhibits better completeness and edge effects.2)In terms of the segmentation speed of physical evidence,a lightweight dual-branch structure of physical evidence segmentation method is proposed.A lightweight dual-branch structure was adopted in the overall network,which including context branches and spatial branches.The lightweight and efficient Mobilenet V2 is used in the context branch as the basic network,and the SE-ASP module is added to improve the feature extraction ability.The spatial information branch is composed of three layers of convolutions to extract the spatial structure characteristics of the evidence image.In order to improve the accuracy of physical evidence segmentation,the weighted multi-scale cross entropy function is constructed for multi-scale monitoring to train the network effectively.Experimental results show that the lightweight segmentation method proposed in the paper is improved by 1.39 times than the Bisenet network,and the accuracy of m PA and m Io U are increased by 1.24%and 0.91% respectively.Meanwhile,the segmentation speed is increased by 9.8 times and m Io U is reduced by 2.4% compared with the improved Deeplabv3+.3)In order to test the applicability of the two methods proposed in the paper in actual scene.The UAV is used to search physical evidence during the experiment and then the images captured by UAV were segmented.The experimental results verify that the improved Deeplabv3+ segmentation method and the lightweight dual-branch segmentation method have advantages in segmentation accuracy and segmentation speed respectively,and solve the problems of poor segmentation results and slow segmentation speeds of other methods.They can respectively meet the demands of two application scenarios,the local processing after data collection,and the on-site synchronization processing.