Research On Image Dehazing Algorithm Based On Sky Segmentation And Convolutional Neural Network

With the popularization of artificial intelligence technology in public life,the application scenarios of computer vision are increasing day by day,and the types of tasks that can be handled are more extensive,such as urban visualization management,remote sensing monitoring,unmanned driving,intelligent security,medical image analysis,etc.have played an important role.However,in recent years,fog and haze weather have occurred frequently in some areas of our country.When light passes through the atmosphere,it is affected by the absorption and scattering effects of suspended particles(such as dust and mist)in the atmosphere.The image quality of the computer is seriously degraded,resulting in a series of problems such as low contrast,color distortion,and image blur,which seriously interfere with the normal operation of subsequent advanced computer vision tasks and affect the processing effect.Therefore,it has a very realistic research significance for image processing systems to implement sharpening processing on foggy images to improve image quality.Therefore,through in-depth analysis of a large number of classic dehazing algorithms,this paper summarizes their respective advantages and disadvantages,and on the basis of further clarifying the reasons for the degradation of hazy images,respectively constructs image dehazing algorithms based on atmospheric scattering model and convolutional neural network.,and the specific research results are as follows:(1)In view of the failure of the traditional dark channel dehazing algorithm for the sky area,resulting in insufficient estimation of the transmittance of the sky area,the restored image is often accompanied by color distortion and halo artifacts,an image combining sky segmentation and weighted fusion is proposed.Dehazing algorithm.Firstly,the mode constraint threshold is set by using the brightness characteristics of the sky area,and the fog image is divided into sky and non-sky areas;secondly,the image edge information is used as the weight map to fuse the dark channel maps of different filter sizes to construct a more advantageous fusion dark channel;then,on the basis of sky segmentation,the atmospheric light values of the sky and non-sky areas are calculated respectively,and the weights are weighted by the proportion of the whole image to obtain a more reliable atmospheric light value;finally,set the transition area to combine the transmittance of sky and non-sky areas.The experimental results show that the proposed algorithm has obvious dehazing effect for fog images containing sky areas,improves the problem of color distortion in sky areas,and suppresses the halo effect in edge areas.(2)The existing dehazing methods fail to make full use of the local feature information of the image,and cannot fully extract the global detail features of the image,resulting in incomplete dehazing and loss of image details.A T-shaped image dehazing network based on wavelet transform and attention mechanism is proposed.Specifically,the proposed network regards discrete wavelet transform as a convolutional layer with fixed parameters,and then obtains multi-scale edge detail features of foggy images by performing multiple discrete wavelet decomposition and reconstruction on the image.The feature attention module,which takes into account the global features of the image and the local information extraction,strengthens the network’s learning in image visual perception and detail texture.Secondly,a T-type connection method is proposed to fuse the extracted image features of different scales step by step,which expands the representation capability of the network.Finally,color balance is performed on the reconstructed haze-free image to obtain the final restoration result.A large number of experimental results in synthetic datasets and real datasets show that the network structure proposed in this paper can fully acquire and acquire image features,and the restoration results are clear and natural,which has better performance than existing network models.