Road Detection Based On UAV Images

UAV images have features such as high resolution and easy access to data,but there are various problems such as road complexity and category diversity in UAV images,using the powerful feature expression capability of generative adversarial networks on target detection,research on how to accurately detect roads on UAV images has become one of the important research directions.In recent years,there has been a lot of research into road information detection on UAV images,due to the fact that different parts of the road in the varying in width and shape,such as village and mountain roads;different types of roads have different color and texture characteristics,such as streets and dirt roads.At the same time,the road area shot by buildings,trees and many other objects blocked,making the accurate detection of road information is still a research difficulty in the field of aerial image detection.In this paper,an UAV image road information detection model is constructed using the conditional generation adversarial network method.The specific research is as follows.(1)Study and improvement of the generative adversarial network model for extracting road information from images of UAVs.In order to address the problem of accuracy of UAV image road extraction results and the shortcomings that the network size is too large to be applied,an adversarial network model based on U-Net idea improves the road detection model of the generator network structure in three main ways: first,it introduces in the network structure down sampling The idea of residuals,through hopping connections can enable feature information to reach deeper into the convolutional network directly to obtain detailed features;secondly,in the bottom layer of down sampling introduces a global pyramid pooling module that can improve segmentation accuracy by aggregating contextual information;thirdly,it introduces a global pyramid pooling module for the To prevent the loss of road information during down sampling,set the minimum size of down sampling feature map to 32×32 when building the network structure.(2)The conditional generation adversarial network is a cyclic adversarial process between the generator and the discriminator,where the generator optimizes its own parameters through the discriminator.A method to optimize the generator by combining the objective function of conditional generation of adversarial networks with the traditional loss function is proposed on this basis model to improve the robustness of the model,and the optimized model can achieve more accurate road detection on UAV images.(3)A comparison experiment is conducted on the standard urban dataset and the PASCAL-VOC2012 dataset to verify the validity of the model.The experimental results show that the model in this paper has good robustness and road detection effect on UAV images.