| The haze disaster has posed severe challenges in areas such as traffic safety,aviation safety,and even military security in China.In these areas,it is important to convert fuzzy images with haze into clear images without haze.Traditional image algorithms mainly perform image enhancement and processing using a priori physical models.Due to the large uncertainty of haze,the effect of traditional image algorithms is not ideal,with problems such as color distortion,blurred target details,and insufficient de-hazing effect.In recent years,with the continuous development of technology in the field of deep learning,the feature of automatic feature extraction by deep learning models provides more selectivity in solving the problem of images with haze.In this paper,under the premise of conducting sufficient research,the image dehazing algorithm is studied,and a deep learning-based dehazing algorithm is proposed for the existing problems of dehazing algorithms.The research in this paper mainly includes the following elements.(1)To address the problems of poor dehazing effect and a small dataset of traditional image algorithms,Cycle GAN is selected as the base network,and an improved Cycle GAN network-based dehazing algorithm is proposed.Cycle GAN network does not need to provide paired datasets,does not need strict correspondence,has wider adaptability,and requires fewer samples.However,the original Cycle GAN network did not achieve the desired clarity and color distortion.In response,modifying the network structure and the loss function was modified to include identity loss to improve the color distortion and lack of dehazing.A channel-level attention mechanism was also added to enhance the network feature extraction capability by adjusting the weight importance of each channel to obtain significantly improve the network performance.The average PSNR of the test set can reach up to 28.3320 and the average SSIM of the test set can reach up to 0.9056.(2)Since Cycle GAN has two generators and two discriminators,the overall network is larger,takes longer to train,is more difficult to train,and requires higher arithmetic power,which is not conducive to practical applications.Therefore,in this paper,the above network is lightened by channel pruning and depth-separable convolution,replacing the traditional convolution with depth-separable convolution and using the lightened NAM attention mechanism to The weights of channel pruning are filtered,and the network structure is fine-tuned during the pruning process to maintain the model accuracy to the maximum extent.(3)The dehazing results are tested on three publicly available datasets,including real and virtual datasets,including indoor and outdoor parts.The results are compared with the screened image enhancement-based dehazing algorithm,atmospheric modelbased dehazing algorithm,and deep learning-based dehazing algorithm to demonstrate the superiority of the two proposed dehazing algorithms. |
