Research And Application On Image Defogging And Enhancement Algorithm For UAV Aerial Images

With the continuous advancement of unmanned aerial vehicle(UAV),UAV aerial photography is widely used in topographic mapping,intelligent transportation,military research and other fields with its advantages of its broad vision,flexible shooting mode and low cost.However,aerial images are easily disturbed by the weather and light,the fog and weak light environment will seriously interfere with the imaging effect of aerial images,and the practicability of images taken by drones in these two types of harsh environments will be greatly reduced.Based on these problems,this thesis carries out the research of defogging and enhancement algorithm based on UAV aerial images.In addition,this thesis applied the defogging and enhancement algorithms to aerial vision tasks in foggy days and nighttime for testing the practical value of the proposed algorithms.The main work of this thesis are as follows:(1)A fast aerial image defogging algorithm called MS-NAOD based on AOD-Net is designed and implemented.MS-NAOD adopts a lightweight multi-scale network structure,this structure can effectively enhance the ability of network to process image details.In addition,a composite loss function including reconstruction loss,structural similarity index loss and total variation loss is designed,and the segmented training method is used to further improve the quality of image defogging.The experimental results show that the MS-NAOD algorithm has satisfactory performance on defogging effect and speed.(2)A multi-scale image fusion enhancement algorithm called MSFEM based on deep learning is designed and implemented.Firstly,based on the camera response model,white balance algorithm and weight fusion algorithm,a high-quality traditional enhancement algorithm WFE is developed.Then,based on WFE algorithm,MSFEM algorithm is designed by combining deep learning and multi-scale structure.Finally,the experimental results on synthetic images and real low illumination aerial images show that MSFEM can effectively enhance low illumination UAV aerial images in real time.(3)The vehicle detection of UAV aerial images in foggy days and nighttime is implemented.In this thesis,by combining MS-NAOD and MSFEM with vehicle detection algorithm,a vehicle detection system for UAV aerial images in foggy and night environment is designed.The test results of the system in fog and night show that the proposed defogging and enhancement algorithm can better assist the vehicle detection algorithm to detect vehicles in aerial images,and verify the practicability of the proposed defogging and enhancement algorithm.This system also provides an idea for UAV to perform aerial photography tasks in complex environment.