Research On Image Dehazing Algorithm Based On Gan Framework And Network Structure Optimization

With the rapid development of industrialization,haze weather has become a common phenomenon.Due to the impact of airborne particles in haze weather,outdoor devices often encounter blurry contours,distorted colors,and reduced contrast when capturing images.This not only affects the visual experience of humans,but also affects the application value of images.Therefore,more and more scholars are devoted to the research of image dehaze tasks.This article focuses on the analysis of existing dehaze algorithms from both the perspectives of common images and remote sensing images.Through the analysis of the advantages and disadvantages of existing dehaze algorithms,two dehaze network models are proposed.Specifically,the following are the contents of the article:(1)Most dehaze algorithms rely on training a network on synthetic paired data,but due to the existence of some gap in the distribution of synthetic data compared to real haze data,this type of dehaze method is limited in practical application.At present,although the dehazing algorithm based on the Cycle GAN network framework can use unpaired data to train the network,the network regards image dehazing as a general image conversion problem,ignores the effectiveness of generator learning,and lack exploration of local regions in the reconstructed images.The only one-level channel attention model used in the network structure is also neglecting the effective utilization of deep channel-related information.To this end,this dissertation proposes a natural image dehazing method based on the improved Cycle GAN framework.The innovation of this method mainly includes the dual-discriminator-based cycle GAN framework and multiple levels of channel attention modules.In particular,the dual-discriminator-based cycle GAN framework adopts the local batch normalization of both generator and constraint generator to improve the convergence effect and increase the attention of local regions.In order to further explore the channel information that is crucial for dehaze image,this article introduces a multiple-level channel attention module based on a one-level and two-level feature statistics,thereby improving the visual quality of dehazed images.The experimental results show that compared to the 8outstanding dehaze algorithms in the public synthetic and real outdoor datasets,the proposed dehaze method achieves the best objective evaluation metrics and visual effect.(2)Currently,most dehaze methods based on U-shaped networks directly transmit encoding layer features to corresponding decoding layers,lacking information exchange between shallow and deep features.In addition,dehaze methods based on non-U-shaped networks have limitation on the perception field,resulting in limited use of context information.As a result,these methods cannot achieve the desired effect in dehazing remote sensing image with large scale changes.Additionally,a single U-shaped network or non-U-shaped network cannot extract more crucial scene details for remote sensing image.Therefore,this article proposes a dual-branch remote sensing image dehaze network that includes level-based feature interaction sub network and multi-scale information extraction sub network.The method uses level-based feature interaction fusion module to integrate semantic information between different levels of shallow features,and spatial details information between deep features,thereby enhancing the information exchange between different levels of encoding layers.Multi-scale information extraction sub network uses multi-scale residual hole convolution module to concatenate different sensing fields of features,thereby obtaining crucial context information for dehazing remote sensing image.In the experiments conducted on two public datasets,the dehaze method proposed in this article achieves the best objective evaluation metrics and visual effect compared to the existing 9outstanding dehaze algorithms.