Research On Image Dehazing Algorithms Based On Deep Learning

Image data is an important foundation for the continuous development of computer vision tasks,and the quality of image data is a key factor that determines the performance of various computer vision tasks.Images captured in the real world are often affected by various adverse weather conditions,and hazy images are one of the common situations.Images taken in hazy weather conditions often show image degradation such as blurring,low contrast,color distortion and other degradation phenomena.Using hazy images as input data for various computer vision tasks(such as object detection,autonomous driving,scene understanding,etc.)will greatly reduce their real performance.Therefore,image dehazing,as a key work in the field of computer vision,has been widely concerned.Especially in recent years,with the development of convolutional neural networks,many remarkable achievements have emerged in image dehazing research.Image dehazing aims to reconstruct the corresponding clear image from a given hazy image,thereby alleviating the impact of image degradation caused by adverse weather environments on various computer vision tasks.Existing single image dehazing methods can be mainly divided into two categories: one is the prior-based image dehazing methods,and the other is the image dehazing methods based on deep learning.Prior-based image dehazing methods design prior knowledge based on the statistical characteristics of the image as constraints for image dehazing process.Although these prior can achieve good results under some environmental conditions,they lack generalization ability for different environmental conditions.With the development of convolutional neural networks and the creation of large-scale datasets,image dehazing methods based on deep learning are emerging,which can directly learn the mapping relationship between hazy images and their corresponding clear images through data-driven approach.In recent years,image dehazing algorithms based on deep learning have made remarkable progress,but there are still some challenges to be solved urgently.For example,how to effectively recover the detail information in the dehazed image to obtain a more natural and realistic clear image;how to improve the robustness of the model while ensuring the dehazing performance,so that it can output consistent dehazing results under different haze density conditions in the same scene.In response to the above problems,this thesis conducts in-depth research from two aspects: how to effectively extract the detailed information of hazy images to improve the dehazing performance of the model and how to improve the robustness of the dehazing model under different haze density scenarios.This thesis proposes a single image dehazing method based on an independent detail-recovery network.Aiming at the problem that the existing dehazing methods ignore the restoration of image detail information after dehazing,this method proposes a detail-recovery network independent of the dehazing backbone network,which considers the dehazing problem and the detail preservation problem of hazy images in parallel.The whole network consists of two parallel sub-networks: one is a dehazing backbone network,which performs a preliminary dehazing operation on the hazy image to generate a coarse dehazed image;the other is a detail recovery network,which extracts local and global detail features from the hazy image,and generates a detail feature map.By guiding with the detail feature map,the coarse dehazed image is refined into a higher quality dehazed image.In addition,this method proposes a multi-faceted loss function,which considers the pixellevel loss,perceptual loss and reconstruction loss of the image simultaneously,and further enhances the stability of the dehazing model.This thesis proposes a robust single image dehazing method based on consistent and contrastassisted reconstruction.This method first studies the robustness of dehazing with different haze densities,and proposes a novel learning framework with different haze densities for single image dehazing.This method introduces a contrast-assisted reconstruction algorithm,which utilizes various negative samples of the dehazed image to compress the dehazed image to its clear target image from different directions,optimizing the traditional positive sample-oriented dehazing objective function.Moreover,this method designs a consistent regularization framework for image dehazing,which ensures that the dehazing model maintains consistent output results for images with different haze densities in the same scene,improving the robustness of the dehazing model.