| The purpose of single image dehazing is to recover hazy images that are taken in the hazy environment into clear images with good visual effects.As an important part of the computer vision field,it has a very important impact on many vision tasks such as target detection and unmanned driving.Therefore,how to improve the single image dehazing effect so as to provide high-quality input for subsequent computer vision tasks has become a hot topic of concern for scholars.Through in-depth analysis of the current research status at home and abroad,aiming at the shortcomings of existing single image dehazing methods,this paper focuses on single image dehazing technology to carry out in-depth research.Firstly,starting from solving the problem of poor image dehazing effect due to insufficient feature extraction in the current single image dehazing method,a single image dehazing method based on two-branch network is researched.After analyzing the advantages and characteristics of smoothed dilated convolution in dense prediction tasks,a feature extraction module based on the smoothed dilated convolution is constructed;Based on the feature extraction module,a two-branch dehazing network with direct learning branch and residual learning branch is constructed,the two branches fully exploit the complementary information between different dehazing models and recover different dehazed images;The two dehazing images are fused by the constructed weight generation module to finally obtain a dehazed image with good visual effects.Secondly,starting from addressing the problem that some detail information is inevitably removed from the original image due to excessive attention to image dehazing resulting in poor visibility of the details of the dehazing results in the current single image dehazing method,a single image dehazing method based on multi-scale features combined with detail recovery is researched.According to the distribution characteristics and imaging principles of haze in images,the multi-scale feature extraction module and multi-scale feature fusion module are designed.The multi-scale features are obtained by using parallel smoothed dilated convolutions with different dilation rates,and the extracted multi-scale features are non-linear weighted fused by applying residual blocks and double attention mechanism;Construct a dehazing network based on the proposed two modules,the preliminary dehazing results are obtained by using this network;A detail recovery network based on image blocking was constructed to extract detail information.The extracted detail information is fused with the preliminary dehazing result to obtain the final clear dehazed image,which can enhance the visual effect of the dehazed image.Finally,subjective and objective valuation methods are used to verify the effectiveness of the two dehazing methods proposed in this paper,ablation experiments and comparative analysis of experiments with existing dehazing methods are carried out on various datasets. |
