Research On Image Dehazing Algorithm Based On End-to-end Neural Networ

When shooting outdoors,it is necessary to dehaze hazy images because haze has a certain impact on the captured image.The traditional dehazing algorithm is to obtain a haze-free image by statistically counting the features of the image through prior knowledge,but in practical applications,when the scene does not conform to the prior,color distortion and detail distortion will occur.In recent years,the rapid development of deep learning has led to significant progress in the field of image dehazing.However,the current method still has many shortcomings,such as the quality of the dehazing image is unstable,and most models mainly rely on the atmospheric scattering model,if the estimation of the parameters in the atmospheric scattering model is inaccurate,the artificial neural network will perform the calculation,resulting in increasing errors of the system,and in the case of dense haze,most algorithms have poor dehazing effect,and the restoration of details and colors is insufficient.In order to solve the above problems,the main research of this thesis is as follows:(1)To solve the problem,algorithms that rely on atmospheric scattering models may suffer inherent performance losses on real images.This chapter proposes a dense connected dehazing network based on cross scale attention feature fusion(DCDNet)based on atmospheric scattering models.DCDNet first processes hazy images using residual-dense blocks,then trains them in multi-scale networks,and then uses upsampling and downsampling to fuse and extract information from different layers at each stage.DCDNet can not only extract more feature information,but also avoid the problem of feature redundancy formed by traditional fully hierarchical multi-scale final fusion.The attention mechanism is used throughout the network to extract more useful information,which can solve the problem of loss of detail caused by simple convolution operations.The experimental results indicate that DCDNet has achieved better dehazing effects on both synthetic and real images.(2)In order to solve the problem of insufficient feature extraction caused by the limitations of the receptive field,firstly,it is proposed that the use of multi-scale module can make full use of the information of the image itself in the hazy image to obtain more effective feature information,and secondly,in order to use the complementarity between multi-scale features to enrich the detailed information in the reconstructed features,an independent feature fusion module is proposed to better capture the context information,which can effectively solve the traditional fusion and lead to redundant feature extraction and reconstruction performance network.(3)In order to solve the problem of difficulty in dehazing caused by the phenomenon of uneven haze distribution,a multi-scale image dehazing network with spatially weighted sparse representation(MSIDNet)is proposed,which is completely independent of atmospheric scattering model.MSIDNet uses the spatially weighted non-local sparse attention module to deal with the problem of uneven haze distribution in hazy images,which can pay more attention to image recovery in dense haze areas.Through extensive experiments,the images processed by MSIDNet are closer to real images than some mainstream existing methods.