| As a more intuitive medium for transmitting information in our society,images often contain more abundant and vivid information.Recently,due to industrialization and urbanization,haze weather has occurred frequently across the country,causing the air pollution index to remain high.In haze weather,a large amount of atmospheric aerosol particles suspend in the air,accumulating a large amount of particulate matter that seriously affects people’s quality of life.The emergence of haze and haze weather makes it impossible for many outdoor devices to obtain accurate image information,resulting in low clarity,low contrast,low image quality,and even image color shift and the loss of many details.This phenomenon seriously hinders the subsequent application of the image.Given this,the purpose of image dehazing is to restore the input hazy image to a clean image.Therefore,studying how to remove haze from blurry images and strengthen details is significant and promising.This article conducts in-depth research and summarizes existing traditional network images and artificial network images,and introduces a deep multi-model fusion network based on contrastive learning.The dehazing performance of haze algorithms on hazy images and our method is fully analyzed through additional experiments and detailed ablation studies.The main research work of this paper is as follows:1.Our proposed improved dehazing algorithm is based on a layered network,addressing issues such as poor interpretability of the convolutional layer in the network,and the estimation of transmission maps and atmospheric light that do not intersect with the dehazing system,thus leading to unsatisfactory dehazing effects.We use the existing dehazing algorithm as our extraction model for the original features of the hazy image,which helps in effectively obtaining features of different levels in the hazy image.We then use the atmospheric scattering model and haze layer separation model to obtain dehazed images,which are fused using an attention mechanism to generate the final dehazed image.Our dehazing network architecture effectively learns long-range dependent feature information,and avoids suboptimal recovery of dehazing results caused by the singleness of the network structure in the dehazing process.As a result,a haze-free image is obtained.The experimental results show that our method effectively addresses these issues,improves the quality of blurry images and outperforms existing technology.2.In deep learning-based dehazing algorithms,most algorithms tend to disregard negative samples(i.e.,hazy images)when training on a dataset.They rely on clear images as positive samples to train the dehazing network.However,experiments show that contrastive learning can help improve image restoration.Therefore,we plan to add the general regularization mechanism based on contrastive learning to the dehazing network,and explore how this affects the algorithm’s performance.By incorporating this mechanism,the image obtained by our dehazing network will be closer to the ground truth and further away from bad samples in the representation space.This will result in a better dehazing image,making it more realistic and natural.3.In this paper,we have proposed an optimized image dehazing algorithm and developed a visual system interface for image dehazing.The user interface of the system consists of three parts: the upload module,dehazing module,and image download module.Through the development and design of the haze removal system interface,we aimed to achieve a simple,convenient,and high-quality haze removal process.By embedding our improved dehazing algorithm into the system,people of all backgrounds can easily use the algorithm for haze removal,and the result has proven its superiority in removing haze. |
