Research On Image Dehazing Methods Via Knowledge Transferring

Haze is a weather phenomenon caused by tiny water droplets or ice crystals which are suspended in the air and close to the ground.Atmospheric light are scattered by these particles,and the visibility of distant objects will be reduced.As a result,pictures taken in hazy days usually suffer from the degradation of color and details,which further affect human’s visual experience and also seriously affect the performance of subsequent visual task such as target detection,target recognition,and scene understanding.In recent years,with the development of deep learning,a large number of image dehazing methods based on deep learning have emerged,and the research on image dehazing has made great progress.However,the existing methods still have the following shortcomings:1)The performance of these methods need to be further improved;2)Most image dehazing methods only give the dehazing results without uncertainty analysis;3)Current methods rely on synthetic datasets and have poor adaptability to real hazy images;4)The exploration of dehazing methods for edge devices with limited computing resources is lacking.At present,knowledge transfer is widely used in image classification tasks and significantly improves the performance of image classification.Inspired by this,this thesis uses knowledge transfer to solve the image dehazing problem.We aim to fully excavate the prior knowledge,and transfer the prior knowledge and learning ability to the image dehazing model by constructing auxiliary learning tasks,and obtain an effective and lightweight image dehazing model.The core innovations and contributions are as follows:(1)A heterogeneous task imitation and knowledge distillation(KDDN)based method is introduced to improve the performance of traditional end-to-end dehazing network.The core of the method is to use the feature representation of the clear image to supervise the reconstruction process of the hazy image.Therefore,KDDN realizes knowledge transferring in the feature domain and improve the performance of the end-to-end dehazing network.Firstly,KDDN trains an autoencoder network to extract the features of clear images,and then treat these features as the dark knowledge learned by the network,which is used to guide the feature extraction process of the dehazing network by featurelevel knowledge transfer.The experimental results show that KDDN achieves significant performance improvement against the contemporaneous state-of-the-art methods,and increases 2.56dB/0.0009 and 0.857dB/0.003 in terms of PSNR and SSIM,respectively,on the SOTS indoor dataset and O-Haze outdoor dataset.(2)An uncertainty-driven image dehazing method(UDN)is proposed to sovle the problem that traditional dehazing methods neglect to analyze the confidence of a model and cannot restore an dehazing image adaptively with the image content.Firstly,UDN uses an uncertainty estimation module to estimate the aleatoric and epistemic uncertainty of the reconstruction results.After that,an uncertainty-aware feature modulation module is proposed to adaptively enhance the learned features.Finally,an uncertainty-driven selfdistillation loss is proposed to improve the reconstruction results by transferring knowledge from high-confidence pixels to uncertain pixels.The experimental results show that UDN can give the confidence value of each pixel while achieving better dehazing performance.In particular,UDN achieves 38.62dB and 0.9909 on the SOTS indoor dataset in terms of PSNR and SSIM,respectively,which is the best result among the contemporaneous state of the art methods.(3)An meta-transfer learning based dehazing method(MetaDN)is introduced to reconstruct various color real hazy image.Unlike the traditional dehazing models which cannot correct color distortion with complicated degradation of hazy image well.This method treats the reconstruction task of hazy images of different colors caused by different atmospheric light as a meta-task,and obtains better initialization model parameters by letting the network learn the ability to process different meta-tasks,so it makes the network to deal with various meta-tasks.Quantitative and qualitative results on various synthetic datasets and real hazy images show that,MetaDN can effectively handle hazy images with different colors.(4)A scalable lightweight dehazing network(SLDN)is proposed to achieve fast,computational resource-adaptive image dehazing.Unlike the traditional dehazing methods with fixed sizes and architectures which have difficulties on being flexibly deployed on different computation platforms.Firstly,this method introduces a lightweight feature extraction module named Hierarchical Ghost Convolution Module to build a lightweight dehazing network.Comparing to traditional convolutional layers,this module has fewer parameters,larger receptive field,and more branches.We can obtain three versions,small,medium and large dehazing models,by adjusting the number of the modules.The three different versions of the dehazing model can be trained in a unified manner,and the optimal version can be adaptively selected for testing according to the computing resources on the device.Finally,an external knowledge distillation mechanism is proposed to transfer the pre-trained knowledge of a wider network to a narrower,lightweight dehazing network to improve its performance.At the same time,an internal knowledge distillation mechanism is introduced to transfer the knowledge of the deeper layers to the shallower layer to improve the performance of the smaller dehazing model.Experimental results show that SLDN can achieve 36.23dB in terms of PSNR on the SOTS indoor dataset with only 1.7M parameters,and the smallest version can achieves 34.15dB in terms of PSNR with only 0.5M parameters.