Exploring Deep Priors In Single Image Dehazing Algorithm

In recent years,more and more computer vision systems are used in all aspects of people's life widely,and computer vision systems have played an important role from daily life to national security.These vision systems are often affected by the deployment environment.Fog is a common bad weather.A large number of particles are suspended in the atmosphere,resulting in serious damage to the images captured by the vision system,such as partial information loss,blurred texture details,color distortion and so on.These problems will directly damage the performance of the vision system.Therefore,quickly and effectively dehazing,improving the image quality and eliminating the interference of haze on the visual system are significant in military security,daily life,and industrial production.Traditional image dehazing algorithms rely on manually designed priors,while the fog distribution in real scenes is often more complex.Therefore,the designed priors have the problem of failure,resulting in poor restoration effect.In recent years,with the successful application of deep learning in computer vision,the image dehazing algorithm based on deep learning has broken through the bottleneck of traditional algorithms and made great progress,but there are still some shortcomings.First,most of the existing methods are designed on the datasets of uniform haze distribution scene,ignoring the problem that the haze distribution is often uneven in the real world,resulting in the serious performance degradation of these dehazing algorithms in practical application;The second is that the existing methods only use positive sample information for supervision,and there is still room for improvement in the restoration effect.Specifically,there are still some problems such as haze residue and color distortion in the restored image,while the potential value of negative sample information has not been mined yet;Third,although the existing deep learning defogging model is constantly improving its performance,it also brings a large number of model parameters,which make deployment more difficulty.This paper explores deep priors in single image dehazing algorithm for alleviating these problems.In this paper,the main contributions and contents as follows:(1)Aiming at the non-homogeneous haze distribution scene,a non-uniform scene image dehazing algorithm based on knowledge distillation is proposed.Considering it is difficult to collect the dataset of non-homogeneous haze distribution scene and the number of existing datasets is few,this paper considers using the model pre-trained by a large number of natural images as the encoder.At the same time,the attention mechanism module is introduced to give different weights to the haze regions with different concentrations,which effectively alleviates the performance degradation caused by uneven haze distribution.In addition,this paper uses clear images to pretrain a teacher network,and migrates the clear images features encoded by teachers to the dehazing network model through knowledge distillation,which further improves the restoration effect of the dehazing network model.(2)To further optimize the dehazing algorithm,a new dehazing paradigm based on contrastive learning is proposed.The key idea of our dehazing paradigm is mining the negative sample information as the lower bound of the solution space.Based on a prior that the output of the dehazing model should close to the positive samples and far away from the negative samples,we propose a plug and play contrastive regularization method.(3)In order to trade off the performance and parameter of the model and mine negative sample information,a compact image dehazing model based on contrast learning is proposed.The model is based on the encoder-decoder architecture.The encoder downsamples the input image,so that the intensive feature extraction calculation is concentrated in the low-resolution feature space,which greatly reduces the amount of parameters and calculation.The feature extraction part adopts an efficient attention mechanism based module and dynamic feature enhancement module.In addition,the model also introduces a dynamic adaptive fusion strategy to compensate for the information loss caused by downsampling.To sum up,this paper takes non-homogeneous haze distribution scene,negative sample information mining,trading off performance and parameter as the starting points to improve and innovate the image dehazing algorithm.The results of extensive experiments demonstrated the effectiveness and superiority of the proposed methods both on the realworld datasets and the synthetic datasets.Besides,the proposed algorithms have been published in CVPR Workshop 2020 and CVPR 2021.