Research On Water Surface Target Detection Method Based On Deep Feature Enhancement In Aerial Images

Rivers and lakes have important resource functions,ecological functions,and economic functions.Therefore,the monitoring of rivers and lakes has received more and more attention from the state,and the identification and positioning of water surface targets is the key task of river and lake monitoring.This thesis will conduct research on water target detection methods and theories,and propose an academically innovative water surface target detection method that can be used in river and lake monitoring.The main work of this thesis can be divided into the following two aspects:1.A water surface target image data set WST2020 composed of aerial images is constructed.There are 6842 water surface target images in the data set,including 5water surface target categories and 23263 labeled target instances.By comparing the WST2020 data set with the general target detection data set and other aerial image data sets,the actual application value and characteristics of the data set are analyzed.Finally,a comparative experiment was designed on the basis of the WST2020 data set to test the performance of the mainstream one-stage target detection model and the two-stage target detection model on the water surface target detection task.The experimental results show that YOLOv3 has the best comprehensive performance in the detection task of water surface targets in aerial images.2.A water surface target detection method YOLOv3-I based on deep feature enhancement is proposed.The method uses YOLOv3 as the basic framework.In the backbone network,through the feature fusion module,the deep features have more shallow details;In terms of multi-scale detection,an improved deep feature learning inception module has been added to the three detection branches.This module can activate the multi-scale receptive field of the deep feature map and add context information of different scales to the deep features.In order to further improve the performance of the method in this thesis,the complete intersection over union loss is introduced in the loss function of YOLOv3-I as the prediction box loss.Experiments show that the method in this thesis improves the detection accuracy of the water surface target.The mean average precision on the WST2020 data set is 77.1%,which is 2%higher than that of YOLOv3.