Research And Application Of Single Image Dehazing Based On Unsupervised Learning

At present,deep learning based methods have become the mainstream method in image dehazing tasks,while most of the research results are focusing on supervised methods.The bottleneck of supervised image dehazing methods is the need to use a large number of paired matched hazed and clear images.Building large-scale paired matching datasets becomes very difficult due to the susceptibility of the scene to natural factors.However,the image dehazing method based on unsupervised learning only needs to pay attention to the hazed image itself,so it is increasingly favored by researchers.In view of the poor adaptability of an unsupervised image restoration framework in image dehazing,the neglect of considering the weight of image features in different spaces and channels,and the large difference between the semantic layout and input of the reconstructed image,an improved unsupervised learning image dehazing algorithm was proposed and related experimental research was carried out.On this basis,a prototype system for image dehazing was designed from the perspective of image dehazing applications.This system can generate clear images by selecting different image dehazing algorithms for input hazed images,which can be applied to corresponding needs.The specific research content includes the following two aspects:Firstly,a single image dehazing algorithm based on unsupervised learning is proposed.This algorithm is an improved dehazing algorithm proposed to address the poor performance of an unsupervised image restoration framework applied to image dehazing problems.This algorithm uses an identity loss to reduce the distribution difference between the dehazed image and the input hazed image,and adds an attention mechanism module to the decoder to assign different weights when processing images in different spaces and channels.The experimental results on the trained network show that the proposed improved algorithm improves by 9.93 and 0.276 on PSNR and SSIM,respectively,compared to the original algorithm,and also outperforms several classic unsupervised image dehazing methods.Secondly,an image dehazing prototype system was designed using Py Qt 5 and Pycharm.The system includes basic functions such as input of hazed images to be processed,selection of image dehazing models,dehazing processing,and output of clear images after processing.This system satisfies users to choose a variety of image dehazing methods based on their actual needs to achieve dehazing processing.