Research On Single Image Dehazing Algorithm Based On Deep Learning

Haze is a common weather phenomenon.Fine particles suspended in a hazy environment can affect the process of reflected light propagation by objects,resulting in quality degradation problems such as reduced sharpness and reduced contrast in the captured images.These problems will seriously affect the effective function of computer vision systems.Therefore,image dehazing has become an important research direction in the field of computer vision.In recent years,single image dehazing algorithms based on deep learning have made significant progress,but there are still some deficiencies.The design of most algorithms relies on the atmospheric scattering model,which will cause the problem of further dehazing effect due to the inaccuracy of intermediate variables;on the other hand,the dehazing performance of the algorithm needs to be improved,and scenes with a large haze concentration cannot obtain a good dehazing effect.Aiming at these problems,this paper studies the single image dehazing algorithm based on deep learning technology.The main contents of this article are as follows:(1)A progressive guidance image dehazing algorithm is proposed.The algorithm is an end-to-end network structure.The network can directly reconstruct the clean image based on the input hazy image,reducing the steps of predicting the relevant intermediate variables through the atmospheric scattering model.Aiming at the problem that the scene with large haze concentration cannot obtain a good dehazing effect,this paper proposes a progressive guidance dehazing algorithm,which realizes gradual dehazing in stages by cascading multiple dehazing blocks.The dehaze operation improves the image dehaze effect.At the same time,adding a guided filtering layer at the end of each dehazing block can eliminate the blockiness and edge blur caused by the dehazing image.In addition,adding depth supervision before and after each guided filtering layer ensures the performance of the dehazing block in the progressive guidance network.Experimental results show that the algorithm can effectively eliminate haze and performs better than existing algorithms.(2)An image dehazing algorithm based on scene-aware cross-channel fusion is proposed.Aiming at the problem that the progressive guidance dehazing algorithm directly takes the final output of the network as a result and has not effectively integrated the intermediate output of multiple dehazing blocks,this paper proposes a scene-aware cross-channel fusion dehazing algorithm.The algorithm can effectively fuse the results of different dehazing stages,allowing the previous stage to focus on the scene with a lower haze density level,and the subsequent stage to learn the scene with a larger haze density level.The dehazing block in the network processes different scene areas separately,and obtains a series of scene-related feature channels.The scene-aware sub-network integrates each channel by learning the confidence of each channel,making the network capable of handling more complex scenes.Experimental results show that the algorithm can effectively integrate the advantages of each dehazing stage and obtain better results than the progressive guidance dehazing algorithm.The research content of this article takes into account both the differences in image scenes and the effect of image dehazing,and has high practicality in real scenes.