Research On Deep Learning Method Of Image Dehazing Under Limited Supervision

Hazy images due to fog and haze often suffer from contrast attenuation,color deviation,imaging blur and information loss,which negatively influence the sequent tasks such as outdoor scene monitoring,video surveillance,navigation tracking and object detection.It is great significant for image dehazing to restore the details and improve the color contrast for the improvement of visual effect.Existing image dehazing methods based on deep learning usually rely on a large number of pairs of hazy/haze-free images,and the collection of paired hazy and clean images in natural scenes is exhausting and time-consuming.Thus,synthetic paired datasets obtained by haze synthesis of clear images are widely used in deep dehazing methods.However,there are domain differences between the synthetic hazy images and the natural hazy images.The dehazing models based on synthetic dataset are poor in performance when facing real scenes.In order to improve the adaptability of dehazing methods and effectively improve the dehazing performance under limited-supervision conditions,limitedsupervision image dehazing is studied and the contributions are three-fold.Firstly,an unpaired image dehazing model with haze prior is proposed.In order to get rid of the dependence of current dehazing methods on paired hazy/haze-free datasets,a dehazing method using unpaired training data is explored.This method adopts the similar framework with Cycle-Consistent Adversarial Networks.Two generators respectively complete the conversion from hazy images to clear images and the reconstruction from clear images to hazy images.In order to effectively guide the network learning and improve the performance of dehazing,the dark channel prior of the original image is used as auxiliary information.Two multi-scale discriminators are used for discriminant learning for the dehazed map and corresponding prior map,so as to integrate multi-scale receptive field and combine various information deriving from the converted image and the prior image.Experiments show that this method achieves good dehazing performance on the general datasets,and shows better generalization performance than traditional dehazing methods and existing unsupervised methods.Secondly,a semi-supervised dehazing model based on consistency training is proposed.Although the unpaired dehazing method breaks away the dependence of synthetic data,the network that only uses unpaired data does not perform well due to the limited learning ability of unlabeled data.Thus,a semi-supervised dehazing model is proposed,which combines the advantages of labeled and unlabeled data to enhance the model’s adaptability in different domains.The proposed method constrains the training of supervised data by minimizing the L1 loss of the dehazed image and the clear image.The dark channel loss and total variational loss are adopted to optimize the dehazing results in the unsupervised domain.Furthermore,the Gaussian mixture model is used to simulate the modalities of haze from the unsupervised hazy samples.The residual distribution similarity of the supervised domain and the unsupervised domain is constrained to narrow the domain gap between the synthetic haze and the real haze.In addition,this method designs multiple auxiliary decoders with different perturbations for consistency training,which effectively improves the performance of the unsupervised branches.The proposed semi-supervised dehazing model shows excellent performance on multiple datasets and can recover better contrast and details.Third,an image dehazing model based on domain adaptation is proposed.The semisupervised dehazing framework based on consistency training uses supervised synthetic dataset and unsupervised real dataset to train the network.Although the adaptability of the model to handle different scenarios is enhanced by learning from real data,only applying the real hazy images for training does not really solve the problem of domain shift.A domain adaption mechanism for image dehazing is proposed to reduce the difference between the synthetic hazy images and the real hazy images,thereby enhancing the model tolerance to different scenes,especially real scenes,and improving the dehazing capability and generalization.Specifically,the CORAL loss is introduced to measure the feature distance between the synthetic domain and the real domain.With the help of domain adaptive feature alignment,the proposed model achieves the superior dehazing performance compared with the state-of-the-art dehazing methods..