Research And Implementation Of Image Dehazing Technology Based On Semi-supervised Learnin

Image dehazing is one of the core research problems in the field of image recovery,which aims to recover haze images to clear images.There are two main types of image dehazing methods,one is the traditional method based on atmospheric scattering model or statistical a priori,which enhances the image by increasing the image contrast to achieve the purpose of image dehazing.However,the method often results in distorted images or a darker overall image,etc.Another utilizes convolutional neural networks to model the image dehazing problem.This type of method uses a large number of pairs of synthetic data to train the model,and although the trained dehazing model has good performance on the synthetic data set,this type of method does not consider the characteristics of inconsistent data distribution between the synthetic haze map and the real haze map,resulting in the model’s less than ideal dehazing effect in the real environment.This paper proposes a new image dehazing method based on semi-supervised learning and meta-assisted learning to address the problem that the current mainstream dehazing algorithms are not ideal for dehazing in real environments.The main research contents of this paper are as follows.(1)We propose a image dehazing method based on semi-supervised learning.The current mainstream dehazing methods use supervised learning to train models on synthetic haze image datasets,and the models trained in this way have poor dehazing effects in real environments.In order to solve this problem,this paper proposes a image dehazing method based on semi-supervised learning,which adds real haze images to the model training process,so that the model can learn more information during the training process and improve the dehazing performance of the model.(2)An image dehazing method incorporating meta-assisted learning is proposed.Based on the semi-supervised learning method,a new branch,called the auxiliary branch,is added to the generator network to address the problem of difficult convergence of the model.The auxiliary branch shares part of the network parameters with the original main branch,and the main branch is trained on the synthetic dataset using a fully supervised approach while the auxiliary branch is trained on the real haze image dataset using a semi-supervised approach.By adding the auxiliary branch,the main branch can better learn the data features on different datasets and speed up the convergence of the model.Through experimental comparison,it is demonstrated that the fused meta-assisted learning has better dehazing performance on the real environment.(3)An image dehazing system is implemented.A system is designed and implemented based on the two image dehazing algorithms proposed in this paper.The system implements the current mainstream dehazing algorithms,including the dark channel dehazing algorithm and the AOD-Net algorithm,and adds the two algorithms proposed in this paper to it.The system also provides a subjective before-and-after dehazing control chart and an objective dehazing effect evaluation index to help users evaluate the merits of dehazing results.