Research On Unsupervised Image Dehazing Based On Improved Generative Adversarial Networks

Image dehazing is a technique to repair blurred images,which can effectively reduce the impact of haze on high-level computer vision tasks.Most of the traditional dehazing methods rely on image processing,atmospheric scattering model,and the prior knowledge of the specific scene.In recent years,deep learning has achieved certain results in the field of image dehazing,but there are also some challenges.Firstly,deep learning relies on a large number of paired images for supervised learning,while it is difficult to collect paired images for training in nature.Moreover,different regions of the hazy image have different noise concentrations,which leads to incomplete defogging in thick fog areas.Secondly,the complexity of deep learning algorithm is high,and the deployment deduction requires more memory and time,which is not suitable for deploying applications.In view of the above problems,this paper studies the image dehazing technology.The main work is as follows:(1)In order not to rely on a large number of paired images,this paper improves the generative adversarial network,designs the DAM-CCGAN algorithm,and uses an unsupervised method for single image dehazing,which can output a haze-free image end-to-end.In order to remove different concentrations of fog in the image more evenly,this paper adopts a dual attention mechanism in the generator.The channel attention mechanism and the spatial attention mechanism are used to effectively focus on the haze in different regions of the image,and different weights are generated for thick and thin fog,which is beneficial to retain a large number of image detail information in the depth fog map.In order to reduce the amount of parameters and calculation of the dehazing model,a 1x1 convolution structure is added to the generator when extracting features,which significantly reduces the complexity of the generator,and a skip connection method is designed in the generator structure to save more image information.In addition,a detail-aware loss function DP＿Loss is designed to improve the image dehazing performance.This paper compares with eight existing defogging algorithms in synthetic outdoor images,synthetic indoor images and real images.In summary,the DAM-CCGAN algorithm is superior to other unsupervised algorithms in PSNR and SSIM indicators,and some index values are better than the supervised algorithm.The complexity is moderate,and the restored clear image is more realistic.(2)In order to make the large model with excellent defogging performance easy to deploy,this paper improves knowledge distillation and applies it to the generator network of DAM-CCGAN,and uses a high-performance teacher model to assist the lightweight student model training for model compression.In the distillation process,the student network uses the attention fusion AF module to fuse the deep features and shallow features,and the multi-level feature fusion helps to obtain more image information quickly.At the same time,attention loss Att＿loss and hierarchical loss Info＿loss are added to strengthen the learning in the distillation process,and the performance of the student network model is trained in many ways.Experiments show that the complexity of the distillation model is significantly reduced,and the trained lightweight generator still has good dehazing performance,which provides a good solution for the application and deployment of the model.