Semi-supervised Learning Based Image Defogging Algorithm

With the rapid development of information technology,human beings have used images as one of the important carriers of information,but in outdoor scenes,the atmosphere absorbs and scatters dust,nitric acid and other turbid media thus forming a haze,in such bad weather,image acquisition equipment is bound to be affected,not only leading to a reduction in the clarity of the captured image,but also causing the loss of image detail information and causing color distortion and other problems.In recent years,computer vision technology has penetrated into various industries,occupying an important position in the fields of aerospace and unmanned vehicles,etc.These systems require high-definition source images as the basis for subsequent image recognition,segmentation and target detection.Therefore,it is of great importance to restore the clarity of images taken under hazy weather.Most traditional image defogging methods are based on atmospheric scattering models or a priori knowledge,but these are not always applicable and the model parameters cannot be accurately predicted,resulting in frequent color distortion and poor defogging results.The advent of deep learning algorithms has enabled networks to automatically learn the model parameters and achieve end-to-end generation of clear images to meet the defogging requirements.However,the limitation of these defogging networks is that they are only trained on fully labeled datasets,resulting in weak model generalization performance,which makes it difficult to apply to fogged images in real scenes.To address the above problems,this paper proposes a new defogging method for fogged images,and the main research includes the following aspects.(1)Firstly,the root causes of image quality degradation under hazy weather are described and classical image defogging algorithms are analyzed.This paper analyzes the main factors affecting image quality from the atmospheric scattering model,combined with the imaging principle of images,and introduces some excellent algorithms previously proposed in the field of defogging,mainly including the DCP algorithm,the AOD-Net algorithm and the MSCNN algorithm.(2)Secondly,to address the problem of weak generalization performance of the fully supervised learning-based defogging algorithm,a semi-supervised learning-based defogging algorithm is proposed in this paper.The network inputs are fogged images and synthetic fogged images in real scenes,respectively.For the training of labeled images,the network extracts the features and projects them into a potential vector and models them with a Gaussian process,and finally uses loss and perceptual loss to jointly compose a supervised loss function and train the labeled data;for the training of unlabeled images,a pseudo-label is generated for them using the previously modeled Gaussian model,and the pseudo-label is used to supervise the training of unlabeled data.training.The proposed defogging algorithm aims to be able to better use unlabeled real image data while utilizing the labeled synthetic image data when the network is trained in order to improve the network generalization performance.(3)The network structure in this paper is similar to the UNet architecture and consists of two parts,the encoder network and the decoder network.The network input is passed through the encoder network for feature extraction,and the extracted image features are mapped into the potential vectors,and secondly,the potential vectors are fed into the decoder network to learn the distribution of the fog in the image to recover the fog-free image.Since Res2 Net has excellent feature extraction capability,each feature extraction block of the encoder and decoder is constructed using Res2 Net,and the defogging network in this paper learns the fog components directly without estimating intermediate parameters to generate a clear image after defogging directly from the fogged image in an end-to-end manner.(4)Finally,the network performance of the proposed defogging algorithm in this paper is analyzed and evaluated through a large number of experiments.Among them,the RESIDE dataset and NTIRE18 dataset are used for the experiments,and the quality of the defogged images is evaluated in terms of both subjective visual evaluation and objective index evaluation.In addition,the performance of the proposed defogging algorithm is compared and analyzed with other defogging algorithms,including DCP algorithm,FFA-Net algorithm,and Dehaze-Net algorithm.Through the experimental data,it is proved that the defogging algorithm in this paper can restore better quality of defogged images.